amal66 makes AI testing a maintainer decision

Port of the amal66/mike fork's Playwright end-to-end suite onto the
upstream backend/ + frontend/ layout:

- e2e/: auth.setup (bootstraps a confirmed e2e@mike.local user via the
  Supabase admin API and saves storageState), auth-flows, critical-path
  (create project -> upload PDF -> ask a question -> streamed response),
  chat-management, project-management, tabular-reviews,
  workflows-account; fixtures/test.pdf
- playwright.config.ts: single-worker (shared test user), setup project
  + chromium project with saved auth state; webServer adapted from the
  fork's monorepo command (npm run dev --workspace apps/web) to
  upstream's layouts: backend `npm run dev` (health-checked on :3001)
  and frontend `npm run dev` (:3000)
- root package.json (upstream has none): @playwright/test, typescript
  dev-deps and test:e2e scripts, trimmed from the fork's root manifest
- root tsconfig.json scoped to e2e/ + playwright.config.ts so
  `npx tsc --noEmit` typechecks the suite
- .gitignore: playwright artifacts + e2e/.auth (session tokens)

Specs select by ARIA role/name and placeholder text only - no
data-testid attributes, so no app-code changes are required.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CEguyEgXa9JjCciXCcVemC

upstream-main (olp's tip) ships the #215/#216 liquid-surfaces UI, so the same
selector drifts fixed on the fork's suite (commit 3d46814 on main) also apply
here. Ports those realigned hunks, plus one upstream-specific route fix.

  * Chat input placeholder is "How can I help?", not
    "Ask a question about your documents..." (ChatInput/TRChatInput). Updated in
    chat-management (rename/delete/project-assistant) and critical-path.
  * The project-assistant empty state replaced the "+ Create New" text link with
    a PillButton reading "Create" (ProjectAssistantTable) - critical-path and
    chat-management now use getByRole("button", { name: "Create" }).
  * The sidebar chat row's active marker is APP_SURFACE_ACTIVE_CLASS
    ("bg-app-surface-active"), not "bg-gray-200/60"; the row wrapper is now h-8,
    not h-9 (SidebarChatItem). The rename/delete tests locate the row accordingly.
  * The documents-toolbar folder button is "Folder" (TabPillButton wired to the
    root createFolderAction in ProjectDocumentsView), not "Add Subfolder"; it
    still opens the autofocused "Folder name" root input.
  * NewTRModal's footer submit and the tabular page CTA both read "Create", so
    the create-review helper scopes to the modal submit
    (button[name="modalAction"][value="create-review"]).
  * The built-in workflow detail test navigated to the flat /workflows/[id],
    which upstream does not route - only the typed /workflows/assistant/[id] and
    /workflows/tabular-review/[id] exist. builtin-cp-checklist is assistant-type,
    so it now navigates to /workflows/assistant/builtin-cp-checklist (the path
    the app itself links to via workflowDetailPath; works on the fork too).

Verified: full 27-test suite green against upstream-main + test-harness +
demo-mode(+provider-registry), backend + local Supabase (upstream schema.sql +
backend/migrations) + MinIO, demo model, e2e@mike.local.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CEguyEgXa9JjCciXCcVemC

Adds .github/workflows/e2e.yml adapted to this repo's backend/ + frontend/
layout: on every pull_request into main (or upstream-main), boot MinIO +
local Supabase (loading backend/schema.sql), start the API and web dev
servers, and run the Playwright suite, uploading the HTML report/traces on
pass or fail. playwright.config.ts already disables its local webServer when
CI=true, so the job owns the stack.

docs/e2e-ci.md documents the one required secret (ANTHROPIC_API_KEY) and the
branch-protection steps that make the 'e2e / playwright' check required, so a
red run blocks the merge button.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Switch the artifact step from !cancelled() to always() so a run that hits the
30-minute job timeout (the failure mode when many specs retry) still uploads
the HTML report and traces instead of skipping the upload - that partial report
is exactly what's needed to diagnose the failing specs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The backend queries exclusively as service_role (lib/supabase.ts, service key,
'bypasses RLS'). schema.sql revokes client grants (anon/authenticated) but
assumes a hosted Supabase where service_role already has full table access - on
a fresh CLI stack loaded via psql it gets none, so the first backend write 500s
with 'permission denied for table user_profiles' and every project/chat spec
fails. Granting service_role after the schema load reproduces the production
grant posture.

Verified in CI: POST /projects 500 -> 201, GET /chat 500 -> 200.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Three fixes that take the suite from ~everything-failing to green (verified end
to end against a local stack):

1. schema.sql lags the migrations (missing e.g. workflow_open_source_submissions,
   which GET /workflows/:id queries → 500). Apply every dated migration on top of
   schema.sql (idempotent; 0 errors on a fresh DB), grant service_role AFTER so
   new tables are covered, then reload PostgREST. Fixes the workflow specs.

2. Serve a production build (next build + next start) instead of next dev. The dev
   server's on-demand compilation (slow first hit → waitForResponse timeouts) and
   hydration-error overlay (injects nextjs__container_errors DOM that pollutes text
   locators, e.g. getByText('All') matching 'Call Stack') made the suite flaky.
   Fixes tabular-reviews list render + the timing flakes.

3. LLM-dependent specs (chat rename/delete/submit, critical-path) require a model
   key to send a message. Guard them with test.skip(!hasLlmKey) (e2e/llm.ts) and
   expose ANTHROPIC_API_KEY to the Playwright process, so a keyless run is green on
   the ~23 other specs and the LLM specs run + enforce only when the secret is set.

Local result (no key, with MinIO+LibreOffice as on ubuntu-latest): 23 passed,
4 skipped, 0 failed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The suite is green with no secret (LLM specs skip). Update docs/e2e-ci.md and the
workflow header to reflect the migrations step, the production build, and the
optional (not required) ANTHROPIC_API_KEY.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The workflows UI refresh (olp/main fa21ac8) renamed the built-in workflow id
builtin-cp-checklist -> builtin-draft-cp-checklist (title 'Draft CP Checklist'
unchanged). The read-only-detail spec navigated to the old id URL and got a
'workflow not found' page. Point it at the new id.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Investigated whether the four LLM-gated specs (chat rename, chat delete,
project-assistant create+submit, critical path) could run keyless with the
demo model they select. They cannot on this codebase: no keyless model
exists here - ModelToggle's MODELS are all Anthropic/Google/OpenAI entries
gated on a configured provider key (env or user-stored,
backend/src/lib/userApiKeys.ts), ChatInput.handleSubmit hard-blocks a send
for an unavailable model, and the backend model set
(backend/src/lib/llm/models.ts ALL_MODELS) has no demo id either. The
'Demo (no key needed)' model the specs target lives only in the amal66
fork. So the blanket test.skip(!hasLlmKey, ...) guard stays as-is - the
evidence does not support narrowing it.

What this change does instead:
- docs/e2e-ci.md: new 'Enable the LLM specs' section with the exact UI
  path and gh CLI command for the ANTHROPIC_API_KEY repository secret, the
  fork-PR secrets caveat (use a maintainer branch or workflow_dispatch to
  exercise the specs), expected per-run cost, and how to confirm in the
  Actions log that the specs ran (27 passed, no skipped line) rather than
  skipped (4 skipped / 23 passed). Also documents the known gap: until the
  specs' selectDemoModel helper is pointed at a Claude model (or the demo
  provider is upstreamed), the secret unskips the specs but they fail at
  model selection.
- e2e/llm.ts: rewrite the doc comment with the precise gate rationale
  (no keyless model in this repo; title-generation 500s are best-effort
  and not the reason) and pointers to the setup docs.
- .github/workflows/e2e.yml: comment-only updates aligning the skip
  narrative (23 keyless specs, not ~20/~23) and pointing at the new docs
  section. No behavioral change anywhere.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

With the directory fetch storm eliminated (previous commit: batched
listProjects?include=documents instead of a getProject burst per modal
open), the local Supabase gateway no longer collapses under modal-open
load, so the specs' 5xx-oriented scaffolding is removed:

- create-with-retry loops (project create x2, chat create x2, workflow
  create, tabular-review create x10-attempt, review add-document x6)
- reload-guards for transient "Project not found" (waitForProjectLoaded,
  gotoProjectRow, project-assistant tab loop)
- networkidle settle-waits whose purpose was to let the per-project fetch
  burst drain before submitting (NewProjectModal / NewTRModal)
- oversized per-test timeouts justified by the storm (180s -> 60-120s)

Kept, deliberately:
- the title-generation settle logic in chat rename/delete
- chat submit's model-seeding retry loop (guards a documented
  useSelectedModel localStorage hydration race, unrelated to the gateway)
- account display-name toPass block (guards a documented profile
  hydration race, unrelated to the gateway)
- generic short networkidle settles after login/logout navigation
- Playwright's own retries: 2 on CI

Verified: with the fix, 3 consecutive full-suite runs (plus 2 more before
de-retry) were green with zero Playwright retries, running against an
account seeded with ~40 projects - a larger directory fan-out than CI's
fresh database ever produces.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The 4 LLM-gated specs (chat-management rename/delete/project-assistant and
the critical-path project flow) selected a "Demo (no key needed)" model in
the ModelToggle. That model exists only in a fork's demo provider - upstream
ModelToggle.MODELS has no such entry - so on this repo, setting the
ANTHROPIC_API_KEY secret unskipped the specs and they then failed at the
model-selection step.

Fix:
- Replace the selectDemoModel helpers with selectClaudeModel, which picks
  "Claude Sonnet 4.6" (the cheapest Anthropic entry in ModelToggle.MODELS).
  With ANTHROPIC_API_KEY exported to the backend, userApiKeys.envApiKey()
  reports the claude provider as configured, so the model is available and
  the submit is not blocked by the ApiKeyMissingModal.
- critical-path Step 8: the canned-reply assertion getByText("Demo mode")
  becomes a presence + nonempty assertion on the assistant answer container
  (MarkdownContent's "div.prose.font-serif.text-gray-900", unique to
  assistant answer content) - deterministic against nondeterministic real
  LLM output, same 60s budget.
- Comments rewritten to describe the upstream reality: specs run only when
  ANTHROPIC_API_KEY is set (e2e/llm.ts), the backend uses the env key, and
  title generation resolves to claude-haiku-4-5 via resolveTitleModel.

Verified against a local prod-build stack (backend :3001, frontend :3002,
local Supabase + MinIO), only ANTHROPIC_API_KEY configured:
- keyless Playwright env: 23 passed / 4 skipped
- ANTHROPIC_API_KEY exported: 27 passed / 0 skipped (1.2m)
- npx tsc --noEmit clean

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The "Known gap" paragraph predicted the 4 LLM specs would fail with the
secret set because they selected the fork-only "Demo (no key needed)"
model. That is now fixed on the e2e suite branch (#220): the helper
selects "Claude Sonnet 4.6" and the critical-path assertion checks for a
nonempty streamed assistant answer. Update the paragraph accordingly:
setting ANTHROPIC_API_KEY now yields 27 passed / 0 skipped (verified
locally; keyless stays 23 passed / 4 skipped).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The two doc commits were written against a lineage where the
selectClaudeModel spec fix lived on a separate branch, which left three
stale claims after the rebase:

- e2e/llm.ts still called selectDemoModel a known gap and pointed at a
  docs section title that no longer exists; the specs in this very
  branch already use selectClaudeModel. Replace the paragraph with the
  current behavior.
- docs/e2e-ci.md attributed the fix to "the e2e suite branch (#220)";
  this branch carries that commit itself, so say so and name the commit.
- The 23/4 keyless and 27/0 with-key figures were presented as bare
  facts ("verified locally"). Attribute them to where they were actually
  measured - the spec-fix commit's own verification against a full local
  stack - and state they have not been re-measured since the rebase onto
  current main. Drop the unattributed "~7 minutes" typical-run figure in
  favor of the structural 27-spec / 4-gated breakdown (confirmed by
  `npx playwright test --list`: 27 tests in 7 files).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>