A pointer-level test and a page-by-page diff against the reference site, rather
than reading the markup.

Dropdowns could be opened but not used. The panel sat 4px below its trigger with
nothing in the gap, and the menu closed the instant the pointer left the
trigger's box — so travelling to an item crossed dead space and the menu shut
before the click landed. The panel now starts flush against the trigger
(`top-full`, padding inside the panel rather than a margin outside it), and the
close is delayed and cancellable. Same contract applied to the language
switcher, which had the identical gap. Verified by dispatching real mouse moves
along the path a hand takes: the menu survives the trip and the click lands on
/en/football.

`columns` silently dropped its children. It reads them from slots col1..colN, so
a child in the generic `default` slot — what a blueprint import or a hand-built
tree naturally produces — matched no column and vanished with no error and no
empty box. That is how the about-us video disappeared. Loose children are now
dealt out across the columns.

Nine of eleven pages had no h1 at all. Blocks hardcoded their heading level, so
whichever section carried a page's title rendered as h2 and the document had no
top-level heading. `heading_level` is now part of the section-header contract
(and of text_image, video, profile_card, contact_form, app_download, cta,
rich_text), so the section carrying the page title is an authoring decision
rather than an accident of which block was used. Every page now has exactly one.

Also from the sweep:
- El3ab's entire privacy policy — nine sections, both languages — was absent
  from our page. Restored from the reference render.
- The line the reference shows above its contact form was missing; contact_form
  gains `form_intro`.
- The chairman's name was a <p> on top of the portrait; it is a heading.
- Footer column titles were <h4> with no <h3> above them, a skipped level in the
  outline screen readers navigate by. Promoted to <h3>.
- `heading_rule` had been added to app_download *inside* its features repeater,
  so every feature row offered a meaningless "underline" control and the section
  heading offered none. Moved to where the heading actually is.

Verified: 22 page/locale combinations answer 200 with no logged block failures,
137 block/variant combinations render, every page has exactly one h1, the
carousel still does not scroll the page, and the suite passes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

fix(website): stop the partner carousel scrolling the page, and turn the hero showcase into a real flipping carousel

Three faults, all found by watching the built site rather than reading it.

The partner carousel dragged the whole page down to itself. Its timer called
`el.scrollIntoView()` every few seconds, and scrollIntoView walks up and scrolls
*every* scrollable ancestor including the document — so a reader anywhere on the
page was hauled to the partners section on a loop. It now sets `scrollLeft` on
its own track, which cannot move anything but itself. Verified: twelve seconds
on the homepage, `window.scrollY` never leaves 0, while the track's own
scrollLeft still advances.

The hero showcase was manual when it should play itself. The reference plays
each design for a beat, turns it on its Y axis to show the back, holds again,
then hands over to the next — an Embla autoplay at `delay: 2500` with a
`duration-700` `rotateY(180deg)` flip and `backfaceVisibility`, read out of
their own bundle rather than guessed. Ours sat still until someone clicked a
swatch. It now runs that cycle: a genuine 3D flip of one object, not a crossfade
between two pictures, with both faces backface-hidden so they never show through
one another mid-turn. Hovering pauses it; reduced-motion skips it entirely.

The colour swatches were large filled circles under the garment, which read as
loud page furniture rather than a control. They are gone by default. A site that
wants them gets `showcase_controls`, and they render as the same discreet
progress dashes the other carousels use.

New fields: `showcase_autoplay` (default on), `showcase_interval` (default
2500ms, matching the reference), `showcase_controls` (default off).

Verified on a local replica carrying the full 11-page site: sampling the hero
across a cycle shows active=1 front → active=1 flipped → active=2 front, so the
flip and the hand-over both really happen; 22 page/locale combinations answer
200 with no logged block failures; 137 block/variant combinations render; suite
passes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Checking the rendered result against the client's own branch data showed the
block was reproducing the wrong rule. Their site lists fifteen branches and dims
nine of them, each captioned with the dates it opens — the dimming tracks
whether the branch is switched on, and the season window is the explanation
shown to the reader, not the test.

Ours filtered `is_active = false` out of the query entirely, so those nine
simply did not exist on the page. A branch under construction is exactly the
thing a marketing site wants to show.

- `getBranches()` takes `$includeInactive`, cached under its own key so the two
  result sets cannot overwrite each other.
- `data_branches` offers `show_inactive`, and treats a branch as dormant when it
  is switched off OR outside its declared season — a branch with no window stays
  open all year, as before.
- Open branches sort first, so an announced-but-closed location never pushes a
  working one below the fold.

Verified on the local replica with two dormant branches alongside seven live
ones: the dormant pair render dimmed with their location and opening dates while
the rest keep the accent border. 22 page/locale combinations answer 200 with no
logged block failures, 137 block/variant combinations render, suite passes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Building an existing client site inside the builder surfaced the gaps between
"the blocks exist" and "the blocks can express a finished design". Each of these
was found by rendering the target and comparing it, not by reading the code.

Blueprint export/import (v2) now carries the whole design, not a third of it.
It exported pages only — so importing a design gave you the content with a
default theme and no navigation, which looks nothing like its source. It now
carries theme settings and menus too. Menu items record their page by SLUG,
because a page id means nothing in another tenant's database and would import a
navigation pointing at whatever happened to hold that id. Tracking identifiers
are deliberately excluded from the whitelist: importing a design must never
start reporting one client's traffic into another's account. v1 files still
import.

Isolation that did not isolate. BlockRenderer catches a failing block so the
rest of the site survives, but Laravel's View::render() calls flushState() when
any view throws, which clears the section stack of the page *around* it.
Rendered inside the layout's @section, one bad block therefore killed the whole
page at @endsection with an unrelated "Cannot end a section" error — the exact
opposite of the intent. Blocks are now rendered before the layout runs, where
there is no open section to corrupt.

Bilingual content reached templates raw. Translatable repeater sub-fields
(a button label, a card title, a partner name) are stored as ['ar'=>…,'en'=>…]
and read straight out of the data array, so they arrived at {{ }} as arrays and
took the block down with "htmlspecialchars(): array given". website_text()
resolves them, and 49 such reads across 15 block views now use it.

The English site rendered right-to-left. website.css hardcoded
`direction: rtl` on .website-body, silently overriding the dir attribute the
layout computes from the locale. Direction now follows the document.

There was no English at all. No lang/ directory existed, so every __() returned
its Arabic key and English visitors read Arabic form labels, buttons and
helper text. lang/en.json covers all 125 public-site strings.

Smaller gaps, each of which made a real design impossible to express:
- 'glass' was a valid navbar template and a forbidden column value; the CHECK
  constraint predated it, so choosing it failed at write time.
- navbar_cta_text had no English twin, so a bilingual site showed one language's
  button to both audiences.
- An empty navbar CTA fell back to the default label, so the button could not
  be turned off.
- An anchor menu item resolved to a bare "#id", which points at nothing from a
  sub-page; it now addresses the homepage in the reader's language.
- A footer column title was a plain string, so it could not be bilingual; the
  'about' column dropped the social row whenever columns were configured.
- product_showcase stacked its showcase under a centred headline instead of
  laying out as the split it is.
- A map field stores coords as ['lat'=>…,'lng'=>…] and the view passed the array
  to urlencode(), killing the block.
- New: 'stacked' info cards, a footer spacer, heading rules and eyebrows on the
  contact block, and a nowrap on highlighted heading fragments so
  "Welcome to {OC-Sport}" cannot break mid-phrase.

Verified against a local Postgres replica carrying a real 11-page bilingual
site: 22 page/locale combinations answer 200 with zero logged block failures,
all 137 block/variant combinations render, reserved and unknown paths still 404,
the four migrations apply and roll back on both an existing tenant and a
from-scratch install + seed, and the suite passes (140 tests, 483 assertions).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The page + block builder shipped in 39f468a9 could not render a single page.
Two independent faults, both fatal:

1. `website/page.blade.php` passes no `$sections`, but `layout.blade.php`
   includes `navbar.blade.php` and `footer.blade.php`, which both dereference
   it. Every builder page died with "Undefined variable $sections" before the
   first block was rendered. `$sections` belongs to the legacy section site;
   both dispatchers now default it and prefer the authored menu tree, so one
   navbar and one footer serve both worlds and existing tenants keep the
   navigation they have.

2. The `ec-*` class layer every block partial styles itself through
   (ec-heading, ec-muted, ec-surface, ec-btn, ec-eyebrow, ec-prose, ec-marquee,
   ec-block) was used in 25+ views and defined in no stylesheet. Even past the
   crash, a rendered page had no colours, no cards, no buttons. The layer is
   now written against the --site-* variables the layout already emits, so a
   theme change repaints the whole site.

Alongside the fix, the capabilities a data-driven marketing site needs:

- Per-language URLs. `/en/...` and `/ar/...` address the same page, SetLocale
  reads the prefix ahead of the session, and the layout emits lang, dir,
  canonical and hreflang alternates from the active locale instead of a
  hardcoded rtl. A shared link now opens in the language it names. `en` and
  `ar` are reserved slugs so a page cannot hide behind a locale prefix.
- One section-header contract: eyebrow, heading, accent rule, subtitle, shared
  by 21 block types through `_header`, with `{braced}` fragments of a heading
  rendered in the accent colour.
- New variants: glass navbar, `overlay_split` profile card, `focus_carousel`
  logo strip, `season_cards` branches. Ambient motion (float/glow/pulse/
  shimmer) exposed in the motion panel; showcase motion on the hero image.
- A footer "powered by" accent band, and footer columns driven by a menu so
  they cannot drift from the navbar.
- Branches gain photo_path and a season window; the block dims a branch that
  is out of season. A branch with no window is open all year, so nothing
  changes for existing data.

Also fixed while in here: `getBranches()` bypasses Eloquent and never checked
`deleted_at`, so a branch deleted in the ERP kept appearing on the public site.

Verified on a local Postgres replica: all 136 block/variant combinations render
with no logged failures, the locale routes answer 200 and reserved paths still
404, migrations apply and roll back cleanly on both an existing tenant and a
from-scratch install + seed, and the suite passes (140 tests, 483 assertions).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The last of the programme, plus the exit gate CLAUDE.md requires before any UI
work counts as finished.

Self-registration (W10) — shipped, and closed
---------------------------------------------
This is the only path in the product that writes into `people` and
`participants` with no member of staff in the loop, so it ships **off**:
`portal.self_registration_enabled` defaults to false on every tenant including
new ones, and the middleware answers 404 — never 403, because a 403 advertises
that there is a signup form here and invites someone to look for the setting.
A public form that starts accepting strangers because a deploy happened is not
a decision anybody made.

Phone verification is the precondition, not a feature. The flow it replaces was
a study in how not to do this: `verify()` accepted the constant '0000' whenever
a seeded setting said 'demo', then resolved *any* active user by phone —
academy owners included — and minted a token with `mobile:*`; and in the other
mode it generated a code, cached it, and never sent it anywhere, so turning the
bypass off locked everyone out rather than securing anything.

So: no bypass exists, in any mode, behind any flag. Only the SHA-256 is stored,
with a bounded attempt count, in a table rather than the cache — a code you
cannot audit is a code you cannot investigate. A send that fails deletes the
record, because a stored code nobody received is precisely the old failure.
There is a test asserting '0000' and '1234' are refused.

Registration goes **through** ParticipantService rather than around it. Writing
the row directly skipped the participant number, the already-a-member check,
the audit columns and ParticipantRegistered — a second creation path that
looked identical and was not. It does not enrol and it does not take money:
EnrollmentService::enroll() needs an actor authorised to enrol and a
self-registering guardian is not one. The member asks; staff enrol.

`people.created_by` is NOT NULL and there is no staff member here, so the
account is created first with no person attached and becomes the author of its
own records — which is also the truth about who typed them. Loosening the
column would have weakened it for every other path.

DuplicateDetectionService runs on every signup and its findings are stored on
the row and shown in the approval queue. It has existed for a long time with
nothing surfacing what it found, so a second Person for an existing member
appeared silently and the two drifted apart forever.

The accessibility gate
----------------------
Eleven checks against the HTML the portal actually renders for a real member,
not against the templates: language matching direction, image alternatives,
accessible names on every icon-only control, a label for every form control,
named landmarks, focus never globally removed, reduced motion honoured, a
stated focus ring, announced errors, and dir=ltr on numeric inputs.

Two things it found. There was no explicit focus-visible style, so the ring was
the browser default — a thin blue line that disappears against a tenant whose
brand is blue; it is now `currentColor`, which inherits an already
contrast-checked colour and is legible on every surface in both themes.
And validation messages were rendered as plain text: a screen-reader user
submitted a form and heard nothing. Every one is a live region now, asserted at
the source, because an error block only renders when there is an error and a
clean page proves nothing either way.

Also: `:focus` gets scroll-margin so a focused control is never left under the
sticky header or the bottom tab bar (2.4.11).

Verification
------------
- 145 migrations from zero on an empty database, seeded, booted a second time:
  every portal grant intact, self-registration absent and therefore closed
  (SettingsService returns the default, which is false — it fails closed).
- The restored oc_sport tenant: nothing to migrate, seeders clean, health 200,
  713 invoices / 356 participants / 649 payments untouched.
- Suite: 76 pass on SQLite; on the tenant PortalSmoke 3/3, AdminScreens 2/2,
  PaymentProof 11/11, CheckInScan 10/10, ServiceRequestEffect 14/14,
  SelfRegistration 13/13, PortalAccessibility 11/11.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Three P1 items, and a latent bug each of them depended on.

SetLocale was registered nowhere. The middleware has existed since early on
and no middleware group ever included it, so `app()->getLocale()` returned the
config default on every request and the bilingual half of an Arabic-first
product was dead code. Worse, that default was 'en' — so every page announced
`lang="en"` while being marked `dir="rtl"`, telling a screen reader two
contradictory things about the same text. The default is now 'ar', the
middleware runs, and the portal's direction follows the locale instead of being
hardcoded.

notification_preferences had per-event, per-channel columns and no interface
anywhere: it was written to by the deleted API and by nothing else, so every
member received everything on every channel with no way to say otherwise. The
preferences screen defaults an unset choice to ON — the member has not asked
for less, and silently defaulting to off means a missed instalment nobody was
told about.

The device list is on the same screen, because push is the one channel whose
recipients a member cannot otherwise see: an old phone, a browser at work, a
device someone else now owns, all receiving silently until the token rotates.

Transfer reconciliation is the control that makes proof approval honest. A
screenshot is not evidence; matching the day's total against the academy's own
statement is. The report is per branch per day with deliberately blank
statement and signature columns, because a report that cannot be signed is not
a control. It also surfaces ageing proofs — a member told "we will check" who
heard nothing — and transfers recorded with no proof behind them, which are
legitimate but which a reconciler needs to expect.

One column asserts a database constraint rather than a number: an approved
proof with no payment is made unrepresentable by
payment_proofs_approved_payment_check, so a non-zero count there means the
constraint is gone, and the screen says exactly that. A constraint nobody ever
looks at is one you find out about the hard way.

notification_preferences.academy_id added to the model's fillable — S1 added
the column and the model was still writing rows that belonged to no tenant.

Verified on the restored tenant: 12 portal screens and 6 staff screens render
200; suite 76 pass on SQLite.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

feat(portal): consent, deletion, requests that take effect, and the seeder bug that would have erased it all

The additions the addendum marked P0 and the programme had not built, plus
two ordering bugs found by testing a from-scratch install rather than only the
incremental one.

The seeder bug (would have broken the portal on the second deploy)
------------------------------------------------------------------
PermissionSeeder deletes every permission_role row for a role and reinserts
its own list, and db:seed runs on EVERY container start when
RUN_SEED_ON_FIRST_DEPLOY is true. So the portal.* grants added by migration
would have worked exactly until the next deploy and then vanished — the
portal 403'ing for every member, with nothing in the logs and no migration to
blame.

And on a brand-new client the migration runs before any academy exists, so it
created no player role and granted nothing at all.

Both are fixed where they belong: the permissions and the `player` role are in
PermissionSeeder and RolesAndPermissionsSeeder now, so a fresh install gets
them and the seeder stops erasing them. The migration stays for existing
tenants. Verified by migrating an empty database from zero, seeding it, then
booting it a second time and re-checking every grant.

attendance.scan reaches trainers, head trainers and reception — the people who
actually stand at a gate. payments.approve_proof reaches accountants.

B2 — consent and deletion (a store-submission blocker)
------------------------------------------------------
Apple 5.1.1(v) and Google both refuse an app that creates accounts and cannot
delete them, so this is what makes a submission possible rather than a
refinement to add later.

There was no consent record anywhere in the schema — not a column — while this
product publishes children's photographs on a public website and sends
marketing over WhatsApp. Consents are versioned and append-only, enforced by a
database trigger: a consent is a statement about a particular text at a
particular moment, so editing one destroys the only thing that makes it
evidence. Withdrawal is a new row. Bumping the document version invalidates
previous answers, because a boolean would silently claim a member agreed to
text they have never seen.

Deletion is redaction, not erasure. This is also an accounting system:
invoices, payments and ledger rows are the academy's books, and a member must
not be able to delete them by tapping a button. The person's identifying data
is destroyed, the login is destroyed, the financial record survives without
their name. Three gates before that: re-authentication, a cooling-off window,
and a blocked state with the reason shown when money is owed or an enrolment
is live — shown up front, because a refusal at the last step is not respectful.

Data export is the other half of the same obligation and is streamed, never
stored: a file of somebody's whole record sitting on disk waiting to be
collected is a second copy of the data they asked to control.

B4 + E7 — requests that actually do something
---------------------------------------------
`grep -rln ServiceRequest` found a model, an event, a listener and a provider,
and no admin screen. Approving a freeze never called ParticipantService::freeze()
— the column changed and the subscription kept running. A member was told
their subscription was frozen when it was not.

Approval is now defined by its effect, and the effect runs in the same
transaction: if it fails the approval fails with it and the request stays
pending. "Approved, and nothing happened" is worse than "still pending".

E7 decided: an excuse is a service_request, never a direct attendance write.
Two contradictory implementations existed and neither worked — ParentExcuseForm
validated, stored its medical attachment to the PUBLIC disk, then discarded the
record behind a `// TODO` while telling the parent it had succeeded; and the
deleted API wrote status='excused' with no marker and no check that the session
belonged to the participant, so a player could excuse himself and corrupt every
attendance figure the product reports. Approval goes through
AttendanceMarkingService with the approving staff member as marker, and a
coach's existing observation is never overwritten. ParentExcuseForm is deleted:
it never worked, so there was nothing to preserve.

B1, B3, B5, B7, B11, B12, B13
------------------------------
- Document upload and renewal. The admin half has been complete for a long
  time — DocumentApprovalList, a nightly documents:expire, a
  MedicalCertificateAlert — and the member half did not exist, so a certificate
  expired at 06:00 and the member had no way inside the product to fix it.
- Payable instalments. reminders:installments and push:installment-due fire
  daily and the only payment path ever built charged the whole due_amount: the
  push said pay and the app could not. Settled from the wallet, which is the
  one payment the portal can complete immediately — money the academy already
  holds.
- Waitlist accept/decline. The offer, the expiry and the push all existed with
  no accept surface anywhere, so the offer expired and the place went to nobody.
- Renewal surface, for RenewalPolicy::ManualRenew, which explicitly means a
  human decides.
- can_authorize_payment gates instalment payment as well as proof submission.
- Every member upload is streamed from the private disk with attachment,
  nosniff and no-store. A member upload is never a URL.
- /health asserts the schema this code needs and names what is missing.
  Verified against a deliberately half-migrated database: 503, and
  ["payment_proofs","invoices.branch_id"].

/parent retired
---------------
Permanent redirects to the equivalent portal screen, parameters preserved so a
bookmarked invoice still lands on that invoice. Two member portals must not
coexist: they diverge, and the one nobody updates is the one a member has
bookmarked.

E2 recorded in config/compliance.php: 18, hardcoded rather than per-academy,
because a settings row nobody tunes is a false choice and a twelve year old
must never open the app and see the household's arrears.

Verification
------------
- 144 migrations from zero on an empty database, then db:seed, then a second
  boot — every grant intact.
- The same on the restored oc_sport tenant: nothing to migrate, seed clean,
  713 invoices / 356 participants / 649 payments untouched.
- 11 portal screens render 200 for a real member; 4 staff screens for an owner;
  members refused on both staff screens.
- Suite: 76 pass on SQLite; on the tenant, PortalSmoke 3/3, AdminScreens 2/2,
  PaymentProof 11/11, CheckInScan 10/10, ServiceRequestEffect 14/14.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

/attendance/{session} binds its parameter to a uuid, so /attendance/scan
registered after it never matched: the wildcard took 'scan' first and died
casting it to a uuid — a 500, not a 404, so it did not look like a routing
problem at all.

Found by rendering the staff screens rather than by trusting that they route:
route:list sorts its output, so it showed the scanner sitting above the
wildcard when the file has it below. The list is not the matcher.

Adds the staff-screen smoke test that caught it, which also asserts a member
account is refused the approval queue and the scanner.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

S6, S7, the staff half of S8, S9 and S10, plus the admin screens S3 and S5
were waiting on.

PWA (S6)
--------
The worker lives at /app/sw.js and is generated per deploy, because the
precache list comes from public/build/manifest.json and the cache name is a
hash of it — a deploy evicts the old cache instead of leaving a worker
serving asset URLs that no longer exist.

Its scope is /app/, not the root: a root worker would control /dashboard and
/api too, serving admins a stale shell and leaving cached credentialed
responses on a shared front-desk tablet.

HTML is never precached. wire:navigate swaps <head> wholesale and prefetches
on hover, so an HTML cache fills with unvisited pages and then injects @vite
hashes from a build that no longer exists — a blank page with no error.
/livewire/* is never cached at all: its snapshot checksum is bound to APP_KEY
and the session, so a replayed one is a corrupt-snapshot error rather than a
stale render. The only offline artifact is a static page with no session and
no CSRF token in it.

nginx gets exact-match locations for /app/sw.js and /app/manifest.webmanifest.
Both end in an extension the static-asset regex claims, and that regex ends in
try_files $uri =404 — so without these the worker 404s before reaching PHP.

Push (S7)
---------
FCM, not VAPID. kreait/firebase-php is installed, device_tokens exists, twelve
listeners already funnel through PushNotificationService, every client has
their own Firebase project, and FCM HTTP v1 delivers to Web Push endpoints
with the same CloudMessage and the same token column. VAPID buys independence
from Google — not a constraint here — for a second sender, table, log path and
prune policy.

So the change is: platform CHECK widened to include 'web', a user_agent column
for sensible pruning, and a unique index on (device_token, user_id) — never on
the token alone, which is what the deleted DeviceController keyed on, letting
anyone claim anyone's token so the victim's phone received the attacker's
notifications. Duplicates are cleared before the index, because a failed
migration blocks every later one on that client forever.

The check-in gate (S8)
----------------------
Staff-scan only. The printed-poster direction stays cut: a printed QR is a
public, permanent, non-secret string, so rotation is impossible by
construction — it proves the member once visited, or knows someone who did.

The scanner screen works with a connected barcode reader by default and uses
BarcodeDetector where the browser has it, because most reception desks have
the reader and not the camera permission.

A real bug the tests caught: participants.status is cast to an enum, so
comparing it to the string 'active' was always false — the gate would have
turned everyone away.

The native shell (S9)
---------------------
flutter_shell/ holds one long-lived Sanctum token in the Keychain or
EncryptedSharedPreferences with the single ability portal:session, and
exchanges it at /app/session-exchange for an ordinary web session in the
WebView's own jar. The token never reaches JavaScript. /app/* is never
exempted from CSRF — that shortcut is what turns a wrapper from safe into
trivially exploitable.

Every bridge is an exported native capability, so each is narrow and checked
natively: the host allowlist is compared against the origin read from the
controller, never from the page; biometrics gate a native action and return
nothing the page can use as an authorisation decision; QR is decoded natively
and only the string crosses.

flutter_inappwebview rather than webview_flutter, because <input type="file">
is inert in a bare Android WebView without onShowFileChooser — and that single
gap breaks the transfer-proof upload, which is the portal's whole money path.

Two endpoints only, and they are the only routes on the sanctum guard. The
deleted API minted tokens with mobile:* — every endpoint it would ever grow.

E3's recommendation stands and the shell is not shipped this cycle. It exists
so that shipping is a decision rather than a project.

App content (S10)
-----------------
One `channel` column on website_news and website_sections instead of the
parallel CMS the plan called for. website_sections, website_news,
website_menus, media, a page builder and website:blueprint export|import all
already exist; a second CMS is a second migration surface, a second editor to
keep in step, and a second place for content to go missing, forever.

Admin screens
-------------
Portal invitations, where the raw link exists for exactly one render and is
never recoverable afterwards. The duplicate-account merge screen E4 asked for
— the prerequisite for ever putting a unique index on users.phone, and the
reason ambiguous phone logins can be refused rather than guessed. Both move
references rather than deleting rows: a deleted user id in a financial record
is worse than a duplicate account.

Documentation that was actively wrong
-------------------------------------
docs/agent-rules/05-financial-integrity.md described double-entry as two rows
with a type of 'debit' or 'credit', and 16-enums-and-checks.md registered that
vocabulary. That schema has never existed — 2024_01_01_000013 created the
single-row shape with both account columns from the start. Anyone writing code
from that text got a mass-assignment no-op and a row that silently said
nothing. CLAUDE.md repeated the same claim, and also said Livewire 3 while
composer.json says ^4.3 — a difference that decides whether a public property
is an IDOR.

The test suite is now symmetrical: tests that build their own tables skip off
SQLite, tests that need a real tenant skip off Postgres, so the whole file
runs clean under either connection instead of one of them being a lie.

Suite: 76 pass on SQLite (24 skipped), and against the restored tenant
PortalSmokeTest 3/3, PaymentProofTest 11/11, CheckInScanTest 10/10.
portal.css is 4.98 kB gzipped against app.css at 31.10 kB.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

S5. The academy publishes a handle, the member transfers and records what
they sent, and staff turn that claim into a payment only after matching it
against the academy's own statement.

An unverified screenshot must never create a Payment. `transactions` are
immutable and recordPayment() forces status = Confirmed and posts to the
ledger immediately, so a proof is a separate object with its own lifecycle
and only approval calls recordPayment() — double-entry happens exactly once
and nothing in the ledger is ever edited.

**A screenshot is not evidence.** It is a convenience. The control is
`sender_reference`, unique per academy per method behind a partial index,
which kills replay, cross-invoice reuse and "someone else's transfer against
my invoice" in one constraint. The reviewer types the amount from the
statement; `amount_claimed` is what the payer said and is never what gets
posted.

Concurrency is a conditional UPDATE, not a disabled button. Two reviewers
open the queue and both see an enabled Approve; the second one's UPDATE
matches zero rows and raises InvalidStatusTransitionException. A row that has
left `pending` is frozen by a BEFORE UPDATE trigger — approving a proof is
the moral equivalent of taking cash, and Auditable::createAuditLog() takes
its user from auth() at boot and silently writes nothing when it cannot
resolve an academy, so the approval facts are columns on the row rather than
an audit-log dependency.

Overpayment is capped at what is due and the excess is deposited to the
member's wallet in the same transaction. InvoiceStatus::Overpaid exists but
nothing consumes it and it drives due_amount negative, after which
getCollectionRate() and ParticipantBillingService start summing negatives.

branch_id is NOT NULL on a proof. Revenue is branch-attributed only through
payments, so a NULL-branch payment lands in the all-branches total and in no
branch — the columns stop summing with no error anywhere.

E6 decided as recommended: all five method CHECKs that lacked `instapay` get
it, the till included. Reception will take an InstaPay transfer within a
month of launch, and the failure mode of leaving the POS out is a Postgres
23514 at the till in front of a customer. pos_transactions and
pos_split_payments also gain `bank_transfer`, which they never had.

The review queue ships before the member-facing upload, on purpose: a proof
that can be submitted and never reviewed is a promise to a member that nobody
is keeping.

Proof files go to the private disk and are streamed by a controller that
authorises the submitter, a co-guardian of the same member, and staff holding
payments.approve_proof — Content-Disposition: attachment, nosniff, no-store.
The parent excuse form wrote its medical attachments to the public disk; that
is the mistake not to repeat.

Verified against a restored copy of backups/oc_sport-20260831-081053.dump: a
600 EGP transfer against a 500 EGP invoice posts 500 to the invoice
(Dr 1010 Bank / Cr 4000 Training Revenue, branch attributed) and 100 to the
wallet; duplicate reference, self-approval, zero amount, second approval and
editing a settled row are all refused.

Suite: 87 passed, 3 skipped locally; 11 InstaPay tests pass against the
restored tenant.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

S3, S4 and the core of S8.

Identity (S3)
-------------
GuardianResolver replaces ten hand-copied
`Guardian::where('person_id', …)->first()` lookups, every one wrong in the
same two ways: `->first()` on a column with no unique constraint, so a
guardian holding two rows saw one set of children and was 403'd on the rest,
silently; and no answer at all for an adult member, because all eleven
app/Livewire/Parent/* components end in ->firstOrFail() and a player has no
guardian row. That is why a player given the `parent` role saw empty lists —
the domain had no path from a user to his own participant.

PermissionService::getChildParticipantIds() also carried
`->where('person_id', …)->orWhere('user_id', …)`, which with the tenant
global scope appended compiles to `person_id = ? OR (user_id = ? AND
academy_id = ?)` — the first branch escaping the tenant filter entirely. The
closure is what keeps both branches inside it.

A real `player` role and the portal.* permissions ship as a guarded
migration, not a seeder: db:seed only runs when RUN_SEED_ON_FIRST_DEPLOY is
true, so a client deployed outside the one-click template would never receive
them. Same pattern as 2026_09_01_000001.

portal_invitations stores only the SHA-256 of its token — a raw token in a
row is a password in a row — and consumption is one conditional UPDATE whose
WHERE clause carries every condition, so two taps on the same link on a phone
cannot both create an account. Activation lives in a plain controller, never
Livewire: a single-use token in a public property is serialised into the page
on every round-trip.

users.email stays NOT NULL UNIQUE, deliberately. 2024_01_01_000002 declares
it inside Schema::create, so Postgres emits a UNIQUE CONSTRAINT that cannot
be made partial without a DROP CONSTRAINT in up(); CREATE INDEX CONCURRENTLY
cannot run in a migration transaction; and password_reset_tokens.email is the
primary key the broker keys on. Portal accounts get p{uuid}@portal.invalid
(RFC 2606, never routable) plus an email_is_synthetic flag every mail path
checks. No unique index on users.phone either: 2026_08_30_000004 logged that
it left duplicates in place, so one would hard-fail on at least one live
client and then block that client's migrations forever.

Phone login now refuses when one number matches several different people —
signing someone into a stranger's account — while still resolving a genuine
duplicate pair for the same person.

config/branch_lock.php gains portal.* and parent.*: RequireBranchSelection
runs on the whole web group, so without it any user holding branches.view_all
in all-branches mode is bounced out of the portal by middleware.

The portal (S4)
---------------
Five tabs at /app — الرئيسية, التدريب, المدفوعات, الأكاديمية, حسابي — with
the pass as a header affordance because it is per active profile: a guardian
with three children needs three.

PortalContext is the scope rule the IA turns on, decided once instead of
eleven times: training is member-scoped, money is family-scoped. The old
components each re-read session('active_child_id') independently while
ParentFinances ignored it and aggregated everyone — the domain saying out
loud that a household has one balance. The active id is re-validated against
GuardianResolver on every read, so a value put into the session, or left
there after a withdrawal, cannot widen what an account sees.

No participant id is held in a public property anywhere in the namespace.
This is Livewire v4, where a plain public property is settable from the
browser, so a check in mount() that is not repeated in render() is
decoration, not a check.

portal.css is the only entrypoint built with `source(none)`. app.css and
website.css are each a bare `@import 'tailwindcss'`, so v4 auto-detects from
the project root and both emit the identical complete utility set — a third
file written the same way would have been a third identical copy. Measured:
portal.css 17.11 kB / 4.54 kB gzipped against app.css at 208 kB / 31 kB.

Screens surface what was always one join away and never loaded: the coach
taking each session and the reason for a substitution, cancelled_reason so an
empty week does not read the same as Eid, and per-event registration for the
right child — answerable only since event_registrations gained participant_id
in S1.

The check-in pass (S8 core)
---------------------------
qr_check_in_enabled has been a toggle in system settings with zero functional
readers since 2026_07_27: the product advertised a feature that did not exist.

The pass asserts identity and never authorizes. Enrolment, participant
status, session existence and branch are fresh reads at every scan, which is
what makes a suspension take effect at the next scan rather than the next
token rotation. The secret is derived by HKDF from a pepper that is
deliberately not APP_KEY, revocation is one integer column, and a scanned
code is consumed by INSERT … ON CONFLICT DO NOTHING inside the same
transaction as the attendance write — a Cache::has/put pair would be a
time-of-check race, and two scanners at one gate is exactly when it loses.
Relay is not solvable; it is made worthless instead.

SelfCheckInService writes through AttendanceMarkingService with the scanning
staff as the marker rather than adding a second attendance write path. The
deleted API had one of those: POST /v1/absences/report wrote status='excused'
with no marker, no transition check, no audit and no check that the session
belonged to the participant.

QrCode is written rather than pulled in — there is no Composer step here that
can add to the committed lock file, and the alternative was the existing
pattern of an <img> pointing at api.qrserver.com, which sends the member's
token to a third party and fails when the venue's wifi does.

It was verified module-for-module against an independent implementation
across versions 1-10 and all eight masks, given identical codewords. That
found two bugs neither visible nor throwing: a Reed-Solomon generator
polynomial built with its terms reversed, and missing version-information
blocks for versions 7 and up, whose 36 modules were being filled with payload
and shifting the whole stream. Both produced a plausible square of black and
white that no scanner accepts. tests/Fixtures/qr_golden.php freezes that
verification.

Verified against a restored copy of backups/oc_sport-20260831-081053.dump:
all seven portal screens render 200 for a real member account, the manifest
is tenant-branded and no-store, and a member opening another family's invoice
gets 403.

Suite: 76 passed, 3 skipped (the tenant smoke test skips off Postgres rather
than pretending SQLite is production).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

S2 of the mobile-portal programme. Branding lived in four uncoordinated
stores with no sync, and each layout resolved it itself in an `@php` block
issuing one SettingsService::get() per field — about sixteen SELECTs per
admin render, repeated on every Livewire round-trip, each with its own
fallback. That is how primary_color came to be defined three times with
three different defaults.

BrandingService returns a readonly BrandProfile, cached under the academy's
new `branding_version` and bumped on save, so it is held until branding
actually changes rather than for a guessed number of minutes, and a queue
worker cannot serve last week's colours. Verified on the restored oc_sport
copy: a second resolve inside one request issues 0 queries.

Defects fixed, each verified against that copy:

- `academies.address` did not exist. AcademySettings has been reading and
  writing it on every save since it was written, and Eloquent silently
  dropped it — no academy has ever had an address stored.
- `branding.academy_name` was read by the parent layout and by every printed
  sheet and written by nothing, so both showed the literal string "الكابتن"
  on every tenant. It is seeded from the academy's own name and is now
  editable. The login page now reads "او سي سبورت" on the verified tenant.
- components/print/sheet.blade.php emitted the raw storage path into an
  <img src>, so the logo was broken on every printed sheet. Paths become
  URLs in BrandingService and nowhere else.
- Guests had no academy bound at all, so the login screen — and the member
  portal's own sign-in, when it exists — rendered under the fallback brand on
  every client. An installation with exactly one academy now resolves it for
  guests too; more than one is ambiguous and binds nothing.
- AcademySettings had no authorize() call while every sibling settings screen
  does.

Dead fields: the plan's rule is wire it or delete it, and none of them
survived as collect-but-ignore. login_background now grounds the login
screen, invoice_header and invoice_footer_text and show_logo_in_invoice
reach the printed invoice, header_bg colours the topbar, compact_sidebar
narrows the rail, and success_color/danger_color colour the flash strip.

Colour derivation. sidebar.blade.php hardcoded `color: #fff` on the brand
accent — this is a tenant-branded product, so a client whose brand is yellow
got white on yellow at 1.53:1. ColorRamp derives a 50…900 OKLCH ramp plus a
foreground chosen by WCAG contrast: that same yellow now gets #111827 at
11.58:1. Nine brand colours are asserted at AA or better.

The ramp is anchored on the tenant's own lightness rather than fixed
targets, because fixed targets are non-monotonic for an inherently light
brand: yellow sits at L 0.86, so a table putting 400 at L 0.70 makes 400
darker than 500. Chroma falls steeply at the pale end — at L 0.97 a chroma
of 0.10 is outside sRGB and clips to mud.

E1 decided as the addendum recommends: `@custom-variant dark` is declared
against the `.dark` class. Roughly 900 `dark:` utilities have been compiling
to prefers-color-scheme and rendering an untested dark ERP for every OS-dark
user, while the toggle did nothing. The OS-driven rendering stops here and
the toggle becomes the only thing that switches themes.

App icons are generated with GD directly rather than by adding
intervention/image: GD is the only image extension in the Dockerfile, and
the whole job is decode, letterbox, resize, write PNG. Dimensions are read
from the header before decoding, since a small file can declare enormous
dimensions. Filenames are content-hashed because nginx serves assets
`expires 1y; immutable`.

SVG uploads are refused everywhere they were accepted. An SVG on the
academy's own origin executes script with the site's privileges and
clean_html() never sees it.

npm run build byte baseline before portal.css exists:
app.css 208.10 kB / 31.02 kB gzip, website.css 217.08 kB / 33.25 kB gzip.

Suite: 67 passed, 0 failed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

S1 of the mobile-portal programme. Every item here is a live defect, and
each one blocks the portal's money path rather than merely preceding it.

The ledger. PaymentService::resolveDebitAccount() returned the literal 1
and resolveCreditAccount() returned 2, both with a `// TODO`. Seeder order
made that Dr Cash / Cr Bank on every payment the product has ever taken —
627 of 701 rows on the restored oc_sport copy — so no revenue account was
ever credited and FinancialOverview::getRevenueBySource(), which groups
transactions by credit_account_id restricted to revenue accounts, could
only ever return []. Accounts now resolve by code within the academy and
hard-fail when absent, and a payment is split across revenue accounts in
proportion to the invoice's own lines, floored with intdiv() and the
remainder on the last row. Routing InstaPay into the old ledger would have
multiplied a broken ledger across a new channel.

The guards. Every rule 05-financial-integrity.md names lived in the UI, in
two hand-copied Livewire components, so any new caller inherited none of
them. amount > 0, amount <= due re-read under lockForUpdate inside the
transaction, invoice not cancelled/paid, academy and currency agreement all
sit in the service now. Draft is deliberately still payable: the POS issues
an invoice as a draft and settles it in the same transaction.

updatePaidAmount() was a read-modify-write on money with no lock — two
settlements landing together each read the old paid_amount and one
increment was lost.

Paymob confirmed callbacks inline: no lock, no Transaction row at all, and
an idempotency guard that was dead code because the finder already filtered
status = Pending, so a retried webhook credited the invoice twice. It goes
through PaymentService::confirmPending() now, which asserts the captured
amount matches.

POS cash sales double-counted the drawer: POSService incremented
total_cash_in and UpdateCashSessionTotals incremented it again, inflating
the expected drawer 2x and producing phantom variance at close. One writer
each now. A split tendered above the total (cash handed over, change given)
capped at the amount due instead of producing an overpaid invoice.

RefundService refunded the full payment only, so an over-approved amount
could not be corrected; it also hardcoded accounts 2/1 with a comment
claiming A/R, which is account 3, and debited the refunding user's own
drawer rather than the one that took the money.

Migrations, all guarded and all verified against a restored copy of
backups/oc_sport-20260831-081053.dump:

- chart of accounts seeded for every academy, not just Academy::first().
  The verified tenant was missing 4060, and db:seed only runs on first
  deploy — so a hard-failing resolver had to be preceded by this.
- invoices.branch_id and transactions.branch_id, backfilled. Revenue was
  branch-attributed only through payments.branch_id, and getCollectionRate()
  scopes invoices through whereHas('payments'), so an invoice with no
  payment yet belonged to no branch. Portal invoices awaiting a proof would
  have vanished from every branch's overdue figure. 588/713 invoices and
  644 transactions attributed.
- academy_id on invoice_items, installments and notification_preferences,
  participant_id on event_registrations — four tenant tables that broke the
  tenancy invariant, all reachable from the portal.
- notification channel CHECK widened to push and whatsapp.
  PushNotificationService writes 'push' and the CHECK allowed only
  in_app|email|sms, so every push delivery log insert raises 23514 today and
  the catch block writes another failing insert.
- guardians and guardian_participant relationship_type CHECKs reconciled to
  their union. NewRegistrationWizard validates one field against the pivot's
  vocabulary and writes it to both tables, so picking أخ / أخت / وصي crashes
  registration on the guardians CHECK right now.
- invoice_number_counters replaces generateNumber()'s count()+1 against a
  UNIQUE(academy_id, number) index — a guaranteed collision the moment
  members can check out without a receptionist serialising them, and it
  reissued numbers soft-deleted invoices still hold.
- the deleted mobile API's INV-MOB lines repaired: it wrote line_total,
  which is not a column, so total_amount defaulted to 0 and every downstream
  allocation read the sale as worthless.

tests/Feature/ExampleTest.php deleted: the stock Laravel scaffold test has
failed since `init` (it GETs / with no tenant database), permanently
red-lighting the suite and masking real failures.

Suite: 61 passed, 0 failed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

AuthOtpController::verify() accepted a constant '0000' in the mode every instance
shipped with, and minted a Sanctum token with ability 'mobile:*' for whichever
active user matched the submitted phone number — staff included. The routes were
deleted in 883391c7, so the tokens reach nothing today, but a credential that was
issuable without authentication should not sit in the table waiting for the next
surface that accepts Sanctum.

Every client gets this, so it is a migration rather than an SSH per instance.
Two such rows exist on the one instance that used the API; the others have none,
and the table guard covers instances that never ran Sanctum's migration.

Deleting rows in up() is a deliberate exception to "destructive operations live
in down() only" — that rule protects schema and client data, and this is neither.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Corrects the rule I wrote two commits ago, which was wrong in the other
direction. I had said rows never belong in a migration. They do, whenever every
client needs them — a lookup table, reference data, a permission the code checks,
a default setting. Seeding those from a migration is the correct pattern, not a
workaround, and add_branches_view_all_permission is the example.

The repository is common ownership: it defines what every client gets. So the
only question worth asking is whether a change is for all clients or for one
specific client. All clients means the repo, and anything touching the database
goes in a migration whether it is schema or data. One client means SSH to that
instance and it never enters the repo — because a migration applies to every
tenant at once and cannot be scoped to one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

I reached for a migration to delete rows — stale API tokens — because the change
needed to apply to every client. That is the wrong test, and the rule it violated
was written down nowhere.

The deciding question is structure or records, never reach. Schema belongs in a
migration because it is structure. Rows belong in the client's own database
because they are that client's data, even when several clients need the same
correction. A migration that edits rows edits them on every tenant at once, with
no review and no way to do it for one client only — and this product is one
install per client, each running the same software over their own records.

Also states plainly that we do not deploy: code is committed and pushed, and the
platform ships it. Nothing here triggers a CapRover build.

Records the code cannot run without — a permission it checks, a default setting
it reads — are genuinely ambiguous, and the repo has precedent both ways. Noted
as a grey area to ask about rather than pretending the line is clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

An expense recorded with a scan attached arrived in the database with no
scan at all. ExpenseForm uploaded the file and passed the path to
ExpenseService::recordExpense(), which builds its Expense::create() array
by hand and never copied the two attachment keys across — so the file
landed on disk and the row forgot about it. It landed on the `public`
disk too, which needs a storage symlink the containers never create, so
even a persisted path would have 404'd.

Receipts now go to the private disk and are read back through
ExpenseAttachmentController, which checks the permission, the academy and
the active branch before streaming a byte.

The list was also a dead end: a row showed a number and a description and
offered nothing but "cancel". Rows are now clickable and carry a view
button, with a paperclip marking the ones that have evidence behind them.

The new detail page is where the expense explains itself — amount,
category, recipient, method, receipt reference, branch, notes, who
recorded it and when, and, if it was cancelled, by whom and why. Below
that sit the journal entries it produced, the original debit/credit pair
and any reversing entry, so the accounting effect is visible rather than
implied. The receipt itself previews inline: images as images, PDFs in a
frame, with download beside them.

An expense recorded without a scan is no longer stuck that way — attach
one from the detail page, replace it (the displaced file is deleted), or
remove it. Every attachment records who uploaded it and when. A cancelled
expense refuses all three: its evidence is frozen with its journal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Another session builds the portal from here, so the entry point has to survive
being read cold.

docs/specs/mobile-portal/ holds three files. 01-program-plan.md is the approved
programme — decisions, workstreams, art direction, the full feature inventory and
verification. 02-critique-addendum.md is a four-lens review of that plan
(completeness, security/abuse, financial integrity, delivery risk) with every
claim checked against the code; where the two disagree the addendum wins, and it
replaces the plan's build order with S0–S10. 00-README.md is the map.

The README leads with four premises the plan was written on that turned out to be
false, because each changes what gets built: Livewire is ^4.3 not 3 (so a public
property is client-settable and validating in mount() is not enough); `dark:`
compiles to prefers-color-scheme with no @custom-variant declared, so ~900
utilities are live and untested rather than inert; `transactions` is one row with
debit and credit account columns, not a pair, contradicting CLAUDE.md and two
agent-rules files; and most of the domain the portal needs already exists.

That last one is the real hazard on this programme. The block-builder engine, the
parent portal, the push stack and the pricing entry points are all built, so the
README lists them explicitly under "do not rebuild" — the plan originally proposed
a second CMS before the review found the first one is generic enough to reuse.

Banners on mobile-app-plan.md, mobile-api-implementation.md and openapi.yaml:
all three describe the /api/v1 surface deleted in the previous commit, and a
native-Flutter-per-client approach that was replaced. Left in place as history,
marked so nobody builds from them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Two live disclosures and one latent account takeover, plus the infrastructure
defects that hid them.

AuthOtpController::verify() accepted a constant '0000' whenever auth_otp_mode was
'demo' — the value every instance was seeded with — and then minted a Sanctum
token for whichever active user matched the submitted phone number, staff
included. It was not exploitable as written, because 2026_08_30_000004 had
normalised users.phone to digits-only local form while normalizePhone() produced
+20…, so the lookup missed. That is one plausible bug-fix away from being live,
which is why the whole surface goes rather than the branch.

Deleting /api/v1 also removes: broadcast/send pushing to every device in the
academy with no permission check; ReceiptController's inverted ownership check,
which made any non-participant invoice world-readable to any token;
PaymentController::initiate with no ownership check at all; and
DeviceController keying updateOrCreate on the FCM token alone, letting one user
claim another's device. None of it is replaced — the member-facing surface is the
session-authenticated web portal, so a second token-authenticated surface meant
building and authorizing everything twice.

bootstrap/app.php built a full diagnostic payload for any 500 and errors/500
rendered it to the browser, ungated by APP_DEBUG. The session it printed carries
password_hash_web — the signed-in user's bcrypt hash — alongside the last ten
queries, the request input and the headers. Now gated on debug, auth keys
stripped by prefix even there, and the production page is self-contained with no
CDN. Detail still reaches storage/logs, keyed by the error id shown to the user.

ParentHome::$activeChildId was validated in mount() and selectChild() but used
raw in render() at eight query sites. Livewire is ^4.3, where a public property
is settable from the browser, so those checks were decoration: a guardian could
walk participant ids and read any child's balance, attendance and evaluations.
Locked, and re-validated in render() since the child list can change between
requests.

ParentExcuseForm wrote the attachment — typically a child's medical note — to the
PUBLIC disk, then discarded the record and flashed success. The parent believed
the absence was excused; nothing was stored, and the record kept feeding the
consecutive-absence threshold that auto-suspends a participant. It now stores
nothing and says so, until excuses are modelled properly.

Infrastructure, because each one hid a failure rather than causing one:
entrypoint continued booting after a failed migration, which serves a stale
schema and silently blocks every later migration forever; the env whitelist had
no PAYMOB_, so config:cache baked null credentials and the gateway failed closed
with no error anywhere; nginx's static-asset regex answered =404 for /sw.js
before PHP saw it; and Route::fallback returned 200 for every unrouted path, so
a deleted endpoint served a website page instead of 404.

Verified: 43/44 tests pass. The one failure is ExampleTest, which fails
identically on unmodified main — confirmed by stashing. Two new tests pin both
disclosures so they cannot return.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The roster's الدفع column answered a different question from the one it
appeared to answer, in three compounding ways.

It showed a LIFETIME subscription total beside a monthly bill. A player who paid
650 in July, 1,200 for a kit bag and 650 in August read as "2,500" for the
current month. Two months of subscriptions were simply added together.

"Subscription" was defined as "an invoice line with no product link" — a
negative definition, so every hand-typed line became subscription money. That
kit bag was typed as free text, so it landed in the subscription figure, was
missing from product revenue, and left the same screen reporting the player had
never bought the kit they had paid for.

The red "has not paid" flag came from an unrelated calculation: matching invoice
text with ilike %اشتراك% plus the programme name. Substring matching on Arabic
also decides that تجهيزي contains زي. On live data the flag and the amount
disagreed on 28 of 247 active enrolments — red rows showing a green figure. The
template's "show unpaid only if flagged AND the amount is zero" guard was not
defensive coding; it was two sources of truth being reconciled where the
disagreement stopped being visible.

The figure is now this billing cycle only, derived from the programme's own
cycle rather than the calendar month, and one computation feeds the amount, the
row flag and the header counts — so they cannot contradict each other again.

Each figure is colour-coded by WHY it is that number, with a legend above the
table: paid in full, pro-rated for a mid-month join, admin discount, line price
override, instalment, partial, unpaid, not yet billed, free. All of it was
already recorded in invoice and line metadata and never surfaced; the reason,
who applied it and the original price now appear on the row. Colour never
carries the meaning alone — each amount also shows a glyph, a label and a
screen-reader sentence, and every case sits at 4.5:1 against white.

The migration links hand-typed product lines to their product where the full
trimmed description matches a product name exactly. Substrings are deliberately
not matched and ambiguous lines are left alone: 44 lines / 209,200 EGP link
safely, 23 lines / 42,800 EGP are reported for a human instead of guessed at.

Verified by replaying the real production rows behind both reported screenshots
through the service: every figure the user questioned is now explained.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Standing authorisation to push fixes without asking was already in place; what
was missing were the checks that make it safe to exercise.

Both failed today. The session-start git snapshot said `main` while a parallel
session had since checked out a feature branch in the same working copy, so a
verified fix was committed to the wrong branch — and `git push origin main`
then reported "Everything up-to-date" and exited 0 while the fix sat elsewhere.
A no-op push is indistinguishable from a successful one unless the remote ref
is checked.

Also makes explicit-path commits mandatory. This checkout is shared with other
sessions whose in-flight work can be staged in the index; `-a` or `git add -A`
would sweep it into a fix commit and deploy it to every tenant.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The financial overview 500'd with SQLSTATE 42P08 on `($4 IS NULL OR
p.branch_id = $4)`. Postgres fixes each prepared-statement parameter's type
during parse analysis, and `:branch_id IS NULL` gives it nothing to work from
— the statement is rejected before it ever reaches the comparison that would
have typed it. `:academy_id` was the same shape and would have failed next.

The idiom came in with 35200985 and could not be caught here: phpunit runs
SQLite in memory, which types placeholders at bind time and executes the
broken form happily.

Fixed by appending the branch and academy filters only when they apply, with
their bindings, rather than passing NULL as a sentinel — which is what the
->when() filters in the same method already do, and keeps the
(academy_id, branch_id) index usable instead of hiding it behind an OR.

The SQL build is extracted to buildTopProgramsQuery() so it can be asserted on
without a database. The test pins four things: the placeholders and the
bindings agree in all four filter combinations, the clauses are omitted rather
than nulled, the built SQL executes, and no raw SQL under app/ binds a
placeholder as a NULL sentinel again. That last one is a source scan on
purpose — the suite's driver is not the production driver, so it cannot
observe this failure by running.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Standing authorisation from the user: a verified fix goes out in the same turn
it is finished, rather than waiting in the working tree for approval.

Written with the order fixed (verify, then commit, then push) and with the
boundary spelled out, because a push here is a deploy to every tenant at once
— entrypoint.sh runs migrate --force and db:seed on every container start, and
there is no staging. Features, schema changes and anything destructive still
get confirmed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

config/branch_lock.php names executive.dashboard as the lock's destination,
but that route was never built. route() throws on an undefined name, so
switching to "كل الفروع" crashed in production after the session had already
been written — the user landed in all-branches mode via an error page.

isLocked() already refused to lock without the route, and that was believed
to make the whole feature dormant. It only made the *gating* dormant: the
guard sits on the decision, while the crash is at the dereference. Four other
sites turned the same name into a URL, and BranchSwitcher's was outside the
gate entirely. Auth/Login reached it only through the config key, so it does
not even contain the string "executive".

Route every caller through BranchContext::redirectRouteName(), which returns
the configured route when it exists and degrades to the dashboard when it
does not. The dashboard is the right fallback while the view is unbuilt: the
lock is dormant, so it is already unfiltered and showing the every-branch
numbers the user asked for.

The test pins the resolver in both directions and fails if any Livewire
component or middleware reads branch_lock.redirect_route directly again.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Taking attendance re-sorted the list by status on every render, so the
moment a coach marked someone the row jumped somewhere else and everyone
below it shifted. Coaches lost their place, could not tell who was already
handled, and recorded the same player several times.

The roster is now ordered by name with the record id as a tie-break —
never by anything the coach can change from this screen — so the list
holds still. Marking a player takes them out of the working list entirely
and into a collapsed "تم تسجيلهم" section, grouped by status with counts,
where the decision can be reviewed or changed. A confirmation toast names
the player and the status that was saved, and a progress card shows how
many are left.

Also here:
- markAs/markPresent/saveRecordNote now resolve the record within this
  session instead of by bare id, and reject statuses outside the four the
  screen offers
- service calls are wrapped in try/catch, so a blocked medical certificate
  shows an Arabic message instead of an error page
- the polymorphic subject relation is eager-loaded with morphWith (was an
  N+1 on every player row)
- one responsive card list replaces the duplicated mobile/desktop markup;
  targets are ≥36px, the progress bar carries progressbar semantics and
  the toast is an aria-live region
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Session::has() is `! is_null(get($key))`, so it reports false for a key
holding null — which is exactly how "all branches" was stored. Three
call sites tested presence that way, so selecting كل الفروع silently
reverted to a single branch on the next navigation and isAllBranches()
was unreachable dead code. All three now use exists().

BranchContext is the one place that reads that state. It lives in
Context, not Services, because the project rule keeps services free of
session/auth so they stay queue-safe; this is the adapter that turns
request state into the explicit ?int $branchId services receive. A null
left by a user whose permission was revoked is repaired rather than
honoured, and stamping deliberately does not follow branchId() — API
routes and queued listeners run outside the request, and a record filed
against no branch would vanish from every per-branch total for good.

The lock itself is dormant on purpose: isLocked() returns false while
the executive dashboard route does not exist, since locking would
otherwise 500 every page including its own redirect target. The
permission ships as a migration as well as a seeder entry, because
db:seed only runs when RUN_SEED_ON_FIRST_DEPLOY is set.

Also stops enabling the query log outside debug — it retained every
statement of every request in production memory for nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

A programme can now bundle products it requires (program_products), so
the group view can answer "who has not bought their registration card"
— which nothing in the system could express before. products.is_essential
is global; this is per-programme.

Each bundled product gets its own column: bought or not, a progress bar,
and the amount settled against the amount billed. Instalments fall out
of this for free rather than needing their own column.

Reading a payment off a line is not possible here — a subscription and a
registration card routinely share one invoice. ParticipantBillingService
allocates each payment across the lines it covers, pro rata on
subtotal_amount, rounding down so the remainder stays unallocated rather
than inventing money. Allocation is capped at the amount billed: a
payment settles total_amount, which also carries tax and fees, so paying
in full would otherwise allocate over 100% of a line. Verified against
production — no invoice over-allocates.

The payment column now shows the amount paid rather than a bare "paid",
with مجاني for free players and لم يدفع for unpaid, and participants
carry their عضو / غير عضو tag. The enrolment-date column is gone.
Total collected is shown to users with invoices.list.

The bundling migration is conditional: it acts only where an academy has
both an active product named قيد and programmes named فريق. Elsewhere it
does nothing, which is what makes it safe for every tenant. On oc-sport
that is exactly one product across 12 programmes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Measured against the live oc-sport database, subscription revenue read
457,970 EGP against a genuine 345,257 — overstated by 32.6% — while the
per-programme breakdown summed to 39,873, about 12% of reality.

Three distinct causes:

POSService::buildInvoiceItems() discarded the item_type/item_id it was
handed, so every POS line landed with a NULL itemable_type. Reporting
reads NULL as "programme subscription", which moved 102,000 EGP of
product sales into subscription revenue — 90% of the error — and meant
no product-ownership check could ever pass. Lines now carry their
Product or Kit. A migration backfills history by matching invoice lines
to their POS lines, filling only NULL rows and only where the match is
unambiguous; a production dry run matched 44 of 45 with 0 ambiguous.

Pro-rata allocation divided by invoices.total_amount, but line totals
sum to subtotal_amount — total_amount also carries discount, tax and
service fees. Every bundled invoice was therefore split on the wrong
denominator (10,713 EGP).

topPrograms joined enrolments to invoices and dropped anything without
an invoice_id. Only 88 of 350 enrolments have one, so 75% of programmes
reported zero. Now a UNION: the exact link where it exists, participant
fallback where it does not, split evenly across a participant's
programmes. Reconciles at 333,105 EGP.

Also: the mounted revenue widgets and the receptionist dashboard omitted
direction='inbound', counting refunds as income, and the widgets' raw
queries bypassed SoftDeletes and cancelled invoices.

EnrollExistingWizard read BasePrice directly, ignoring membership type
and every pricing rule, so it quoted a different figure than the
registration wizard for the same player. Both now go through
PricingService.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Section and theme editors wrote straight to columns that carry CHECK
constraints, so a bad value surfaced as a 500 rather than a field error.
Adds rules() mirroring the constraints, Arabic messages(), and an error
summary in both forms.

home() also served unpublished sites to the public. Staff keep their
preview route; everyone else is sent to login.

Drops a redundant invalidateAll() from ThemeEditor::save(): the call
passed an argument the method does not take, and would have flushed
every tenant's cache. WebsiteSettingService::update() already
invalidates the one academy that changed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Records were reaching the database with no branch, so they belonged to
no branch and were invisible in every branch view. Three causes:

1. Invoices have no branch_id column, yet three call sites read
   $invoice->branch_id and stored the result. It was always null.
   POSService did this for every point-of-sale payment, which is why the
   walk-in ("عميل عابر") sales had no branch. POS now uses the branch the
   sale was rung up in; the mobile payment controller and InvoiceShow
   take it from the participant being billed.

2. PaymentService::record() only set a branch if its caller happened to
   pass one, and most callers did not.

3. Nothing enforced the rule centrally.

New BelongsToBranch trait stamps the active branch at creation, mirroring
BelongsToAcademy. It is applied to the models that record an action —
Payment, Expense, CashSession, FacilityRentPayment, POSTransaction,
PurchaseOrder, Participant, TrainingGroup — and deliberately not to
catalogue models such as BasePrice, PricingRule, Product and Employee,
where a null branch legitimately means "shared across all branches".

The trait adds no global scope on purpose: branch is a reporting lens,
not an isolation boundary, and scoping globally would break console
commands, cross-branch reports and the switcher's "all branches" mode.
It also returns null rather than guessing when there is no request
context, so scheduled jobs do not misfile academy-wide records.

Also adds a migration trimming stray whitespace — including the
non-breaking space U+00A0 that survives copy-paste — from names shown to
users. Those characters are invisible in forms but render as a gap in
page titles and receipts, and break exact-match lookups.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Three separate defects made the financial figures wrong.

1. Refunds were counted as revenue. Eighteen queries summed payments on
   status='confirmed' with no direction filter, so outbound refunds were
   added to income across the dashboard, the revenue/product/subscription
   widgets, the financial report, the print report and ReportService.
   That inflated revenue by 40,048 EGP all-time, 32,510 this month.

2. Refunds were simultaneously counted as an expense. The refunded
   original already drops out of revenue when its status becomes
   'refunded', so adding the outbound payment to expenses deducted the
   same money a second time. Refunds are now contra-revenue: the revenue
   card shows gross collected, refunds, and the net, and the expense side
   no longer includes them.

3. Expenses were presented as vague lumps, the worst being "مدفوعات أخرى"
   — which was in fact customer refunds. The breakdown is now one line
   per real cost (payroll, facility rent, purchases, and each expense
   category separately), sorted by size, each stating where it comes
   from.

Payroll was missing from expenses entirely; approved and paid payslips
plus trainer compensation are now included, scoped by branch through the
trainer's employee record.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The financial overview filtered revenue by branch but not expenses, so
every branch showed the same expense figure. The "مدفوعات أخرى" line was
academy-wide outbound payments — 32,510 EGP of customer refunds issued
this month, all belonging to Zayed — displayed identically under all 7
branches. Purchase orders had the same problem, and in the 6-month P&L
chart both income and outbound were unfiltered.

Also: PaymentService::refund() created the outbound payment without
copying branch_id from the payment being refunded, so refunds taken
through that path landed in no branch at all and were invisible in every
branch view. RefundService already did this correctly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

102 KB of planning documents (website-builder-v2 spec, mobile app plan, mobile
API implementation, data snapshots) were sitting directly in .claude/ rather
than in docs/. They are reference material, not agent configuration.

.claude/ is now empty of markdown, so nothing in this repo is auto-loaded into
agent context except CLAUDE.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The 18 files in .claude/rules were injected into every agent turn — 38 KB of
POS, inventory and attendance rules loaded even while editing CSS. They are now
in docs/agent-rules/ and read on demand.

CLAUDE.md keeps every hard invariant inline (money as piasters, tenancy scoping,
double-entry immutability, migration-first, RTL logical properties, no dead
links, safe_url/clean_html) and indexes the detail, so nothing that protects
code quality was dropped.

Also documents the deployment constraint that governs every migration: all
tenants build from main and the entrypoint runs migrate --force plus db:seed on
every container start.

Net: ~57 KB less agent context per turn.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Captured 2026-08-30 from oc-sport.com as the source spec for rebuilding the
client's site in the v3 builder. Contains their 7 pages in both locales, the
static assets, the public images, their published OpenAPI description and the
extracted bilingual content catalogue (604 strings x en/ar).

analysis/FINDINGS.md holds the gap analysis that motivated the v3 builder.

Excluded from the Docker build context via .dockerignore, so it never ships to
a tenant image.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Menu items, button links, announcement bars, popups, floating buttons and
navbar CTAs all wrote editor-supplied values straight into href attributes.
Any user with settings.manage could store a `javascript:` URL and have it run
for every visitor of that tenant's public site.

Adds safe_url(), which allow-lists http/https/mailto/tel/whatsapp, site-relative
paths and fragments, and rejects protocol-relative URLs, data:, vbscript: and
entity/whitespace/control-character obfuscation before testing the scheme.

Applied at four layers so no writer can bypass it:
  - input     MenuManager rules + BlockField Link validation
  - model     WebsiteMenuItem::href()
  - render    every editor-supplied href in every website view
  - import    WebsiteBlueprintService, since blueprint files skip form rules

safe_url() returns null rather than '#', so blocks skip the link entirely
instead of emitting a dead anchor — this satisfies the project's no-href="#"
rule with the same mechanism.

Also fixes Alpine expression injection in the gallery and schedule lightboxes,
where a quote inside an image URL could break out of the inline handler: Blade
escapes ' to ' but the browser decodes it before Alpine parses.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The v2 builder could not express more than one page: website_sections had a
unique(academy_id, section_key) constraint, there was no pages table, and
SectionManager exposed only toggle + reorder. A client with a seven-page site
had no way to represent page two.

Adds an additive page/block model alongside v2:

- website_pages + website_blocks (nested tree, JSONB data/style)
- BlockRegistry of BlockType classes: 31 types, 133 layout variants, 237
  fields, 442 validation rules derived from the field schema
- Page/Block/Menu/Blueprint services, BlockRenderer, BlockDataResolver
- Builder UI: page manager, block tree editor, schema-driven field forms,
  repeaters, content/design/motion panels, image upload
- Authored navigation (website_menus) with dropdowns, replacing nav links
  that were previously derived from enabled sections
- Blueprint import/export via `php artisan website:blueprint`
- Extended motion library: entrance effects, delay, stagger, parallax

A new block type now costs one PHP class — no migration, no enum case, no
CHECK constraint.

Nothing here is destructive. website_sections is untouched and "/" falls back
to the legacy renderer when no builder homepage exists, so already-deployed
tenants are unaffected until they opt in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The enrolment wizard looked up a programme's base price without any
branch filter, so with per-branch pricing across 7 branches the
receptionist could be quoted another branch's price.

The registration wizard's printed receipt resolved the branch with
Branch::first(), so every receipt printed the first branch's details
regardless of where the registration actually happened.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The system predates branches, and adoption of the branch switcher was
partial: 44 of 186 Livewire components used UsesBranchScope, and several
that imported it never actually called it. The dashboard was the worst
case — half its widgets were branch-aware and half silently reported
academy-wide totals next to them, so the numbers on one screen were not
comparable with each other.

OC-Sport runs 7 active branches, so every unscoped widget was showing
six other branches' data.

Dashboard: scoped trainers-present, pending payslips, pending documents,
low stock and expiring medical certificates, which were academy-wide.
All six dashboard widgets (revenue, product revenue, subscription
revenue, enrolment trends, overdue renewals, trainer dues) now filter by
branch, including the raw-SQL CTEs in the revenue breakdowns.

Lists and reports: events, evaluations, base prices, pricing rules,
promotions, stock counts, kits, document approvals, trainers, trainer
advances, payroll, essential deliveries and the financial report.

Pickers: participant, group, program, facility, warehouse, product and
employee selectors now offer only the active branch's records, so a
transfer or invoice cannot silently reference another branch.

POS and InvoiceShow used auth()->user()->branch_id directly, ignoring
the switcher entirely — a user who switched branch still transacted
against their home branch. Both now read the active branch.

Deliberately left unscoped: parent- and guardian-facing screens, which
are scoped to their own children and have no branch switcher, and
single-record detail screens, which are already scoped by the record and
would hide legitimately related history for participants who moved
between branches.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>