- test_mock_response.py: split the module docstring summary line (D205 was
  failing pre-commit.ci, the only red hook on the PR).
- groups.yaml: drop `optional: true` from e2e-workflow / e2e-api-deployment.
  They now hold real execute-path tests rather than placeholders, so a red one
  should fail the build directly instead of only surfacing as a critical-path
  gap via --fail-on-critical-gap.
- README: the workflows/ and api_deployment/ dirs are no longer "(future)", and
  document UNSTRACT_LLM_MOCK_RESPONSE under E2E runtime — including that only
  completions are mocked (embedding.py is unhooked; chunk_size=0 is what keeps
  indexing off the real provider).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P4tU5vyB4G6BgfAnGqjxLs

The comment block sat where #2176 reworded the e2e-login comment and inserted
a new group, creating a textual merge conflict against main. Move it inside
each group body; the optional: true removals merge cleanly on their own.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P4tU5vyB4G6BgfAnGqjxLs

Drop the restated-critical-path and mechanism-narrating comments across the
LLM mock and the execute-path e2e tests; keep only the constraints the code
can't show (litellm's fixed usage, the mandatory filter arg, the two-call
execute contract). Also drop the groups.yaml 'not optional' comment: the field
is self-evident and would go stale.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P4tU5vyB4G6BgfAnGqjxLs

Three follow-ups from review:

- llm.py: warn once per process when UNSTRACT_LLM_MOCK_RESPONSE is active. The
  hatch was silent, so a stray env var in production would fake every
  completion and its billing with nothing in the logs.

- rig: implement the runtime-gate-skip TODO in cmd_run. Groups run in topo
  order, so a dep's result is known before its dependents start; a red dep
  means the precondition it gates does not hold and dependents would only
  produce noise against a half-up stack. Adds GroupManifest.transitive_deps and
  a 'blocked' GroupResult status which is never green, so a blocked group
  attests no coverage and --fail-on-critical-gap still catches a path that
  quietly stopped being covered. This mattered less while every platform
  dependent was optional; it does now that the execute-path groups gate.

- api-deployment e2e: the 10/20/30 usage assert now explains itself, since
  counts are summed per reason and a multiple means extra completions rather
  than broken tracking.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P4tU5vyB4G6BgfAnGqjxLs

Closes the remaining gaps in the registry; every path is now proof: marker.

- callback-result-delivery: async API execution (timeout=0). With a file to
  process, COMPLETED is written only by the chord callback, so a polled result
  proves the callback worker ran. Pins the one-shot read too — fetching a
  delivered result acknowledges it and the next read 406s, which a retry loop
  added later would otherwise swallow silently.

- workflow-execution-fan-out: multi-file execution, one batch per file. Needed
  MAX_PARALLEL_FILE_BATCHES on the *backend* (workers ask it for the value and
  only fall back to their own env); at the default of 1 the files share a batch
  and run serially, so the test would have passed proving nothing. File bodies
  differ because identical ones are deduplicated by hash on ingest.

- prompt-studio-fetch-response: the authoring surface. Async, so the answer is
  polled from prompt-output/ rather than the task status, which reports the
  executor finishing and not the callback having stored anything. Upload is
  content-sniffed and OSS takes PDF only, hence the real (if empty) PDF.

- pipeline-etl-execute: source connector to destination, over MinIO — the only
  storage connector the compose stack both boots and registers. Reuses the
  exported tool but needs its own workflow, since endpoints are per-workflow
  and the existing one is committed to the API deployment tests.

None of these have run: the e2e job is skipped while the PR is a draft. Locally
verified only as far as it goes here — rig validate, rig self-tests, ruff, and
pytest collection.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P4tU5vyB4G6BgfAnGqjxLs

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017UVfw7aAocC3KCJaEyZ5GU

Rig:
- Blocked groups now persist a synthetic junit carrying `rig-blocked-by`, so
  the cascade survives `rig report`, which rebuilds results from junit alone.
  Without it a blocked group parsed as a zero-count pass and falsely attested
  critical-path coverage.
- GroupResult rejects a blocked row with non-zero counters.
- Report statuses are module constants instead of scattered literals.
- An exit-5 (no tests collected) group that others depend on now counts as
  failed, so dependents are blocked rather than run against nothing.
- The LLM mock default moves to cmd_run so it applies to every runtime.
- groups.py docstring no longer claims runtime gating is unimplemented.

E2E:
- Shared `wait_for_execution` helper replaces three near-identical polls; a
  CsrfSession subclass stamps X-CSRFToken so per-call helpers can't drift.
- api_deployment poll: sleeps on every iteration, captures exceptions as
  evidence, treats 406 as a lost-200 acknowledge rather than a poll assert,
  and surfaces non-422 statuses as themselves. Timeout drops to 150s so the
  group's three tests fit one 600s budget.
- MinIO scheme drives both the workers' URL and the client's TLS flag.
- Missing rig-provisioned prerequisites fail instead of silently skipping.

Core/SDK:
- `ExecutionStatus.terminal_statuses()` names the terminal set.
- Mock-response tests cover the empty-string no-op and assert every
  completion path still calls the injection hook.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017UVfw7aAocC3KCJaEyZ5GU

The e2e image build is the only CI check that runs `bun install` and
`bun run build`; ci-frontend-lint only runs biome. Excluding frontend/**
from e2e_relevant let a broken build or a desynced bun.lock merge green
and surface on main via production-build.yaml.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017UVfw7aAocC3KCJaEyZ5GU

The fan-out test asserted only the rejoin: every assertion held identically
for three files processed serially in one batch, while critical_paths.yaml
retired the workflow-execution-fan-out gap on its strength.

Nothing distinguishes a batch — the batch index is a discarded loop local and
the celery task id reaches only worker stdout — so assert the timing instead.
WorkflowFileExecution rows are created by the worker owning the batch rather
than up front by the dispatcher, so fanned out each row's window is one delay
wide, while serialised into one batch the windows open together and grow by a
delay per file. Comparing durations rather than overlap is the point:
serialised rows are created in one tight loop and overlap too.

That needs each completion to cost a known, measurable amount, so add
UNSTRACT_LLM_MOCK_DELAY alongside the existing mock hatch. It is inert without
UNSTRACT_LLM_MOCK_RESPONSE, since litellm ignores mock_delay on a real
completion, and a non-numeric value degrades to no delay rather than breaking
every completion in the worker.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017UVfw7aAocC3KCJaEyZ5GU

Manual runs reported green with nothing tested. paths-filter has no base to
diff on workflow_dispatch, so it resolves against the default branch and a
manual run on main diffs main against itself: every filter false, every tier
skipped, gate green. Filtering is a PR-time saving, so force both scopes true
off-PR and skip the filter step entirely there.

Gate the LLM mock on ENVIRONMENT as well as the mock var, so a stray variable
in a real deployment can't fake completions and their billing. Production sets
neither, so this fails closed. `development` is allowed alongside `test`
because that is what base compose sets on the two workers that run the
injection -- proven by the mock working today with the var set on exactly
those two services -- and the overlay now pins ENVIRONMENT=test on them so the
tier can't lose its mock to a base-compose edit. A refusal warns rather than
passing silently: setting the var at all means someone expected mocking.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017UVfw7aAocC3KCJaEyZ5GU

The rig's `down -v` destroys the containers, so the workflow step that ran
`docker compose logs` afterwards always wrote an empty file — and overwrote
nothing useful only because there was nothing else to write. Capture inside
`ComposeRuntime.down()` while the containers still exist, and drop the step.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The bound was half a delay — a 1s window on per-file durations of ~12s, which
measures runner contention as much as batching. One delay is the midpoint
between the two outcomes the check separates (~0 spread fanned out, ~2 delays
serialised), and the delay itself has to be large enough for that midpoint to
sit above the noise.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The worker log settles what the test could not: `Arranging 3 files into 3
batches` — the fan-out happens. Yet the per-file durations it was judged by came
out 6.6 / 18.1 / 22.9. Three tool containers running at once on a CI runner cost
more in contention than they gain in overlap, so genuine fan-out lands further
apart than serialised files do, and no threshold separates the two. Raising the
delay does not help: the contention scales with the parallelism being measured.

So the proof is removed rather than retuned, and with it the mock delay that
existed only to feed it -- every mocked completion across the whole e2e tier was
stalling for it. `workflow-execution-fan-out` becomes a declared gap (visible in
the report, non-gating), and the test keeps asserting the rejoin, which is
directly observable. Closing the gap wants a batch id on WorkflowFileExecution,
which is execution observability worth having on its own.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>