Whole-video questions, the fold that makes them affordable, and a container to run it all in

[vimal-55]

Three separable changes, one per commit — read it commit by commit.

1. Resolve a locally ingested video to the file in samples/

local_file knew only the make sample-real layout (<video_id>_<youtube_id>.<ext>). A locally ingested video lands as <video_id>.<ext> and never matched. It indexed, answered and cited fine, then the frontend drew "no playable copy" over a video sitting on disk. Both globs keep the separator explicit so 61 cannot match 611_H8fGd3fCJbg.mp4.

2. Answer whole-video questions, and fold the transcript so the tier can

"What is this video about?" has no answer in any one 25-second chunk, so the extractive path declines it — correctly, under answer_v1.md rules 2 and 3. The synthesis happens once per video into runs/<stem>/overview.json, and every claim carries seconds off a real chunk so the existing ground() and the existing player handle it unchanged.

The 413 is the finding in this PR. The first build produced a document whose people held "Baroque period", "Victorian England" and "Rome", every span crammed into the last 300s of a 1805s video. That was not the prompt — it had run on the 3B local model, because the groq arm fell back:

413  {'code': 'rate_limit_exceeded', 'type': 'tokens'}
"Request too large ... tokens per minute (TPM): Limit 8000, Requested 17152"
x-ratelimit-remaining-tokens: 8000      <- a FULL bucket
x-should-retry: false

Groq reports both capacity failures under rate_limit_exceeded, so matching the string read a permanent condition as throttling. The bucket was never empty; the request cannot fit. A 413 now never falls back, whatever answer.fallback says.

No real transcript fits one call on this tier, so build windows the transcript and merges. people and topics are merged in code, never by a model — every span has to be a span off a real chunk, and concatenation cannot invent one. Only the abstract needs synthesis, and it is given no timestamps at all.

Three levers, each found by a failed run rather than estimated:

lever	was	now	why
overview.max_context_chars	180000	10000	180000 measured the model's context window; the binding limit is the tier's throughput. Measured 3.531 chars/token, 2528 chars fixed overhead.
overview.window_max_tokens	—	4000	A window is a quarter of a video but the cap is charged whether used or not. At 2500 the reply was cut off before topics.
answer.reasoning_effort	—	"low"	gpt-oss reasoning tokens are charged against the completion cap. The failing window finished in 1539 tokens at "low".

3. Run the pipeline, the gates and the API in a container

A toolbox image, not a server: a supervisor re-runs the gate command and compares output, and an image that could only serve /ask could not be handed to one. Ollama is a sibling service because src/embed.py and src/retrieve.py call ollama.embed and there is no hosted embedding arm.

docker/ollama-init.sh pulls the :F16 tag written out in full — an untagged pull takes Q2_K and config.toml records what that costs (recall@5 0.9167 → 0.4167) — and refuses to run if that tag has drifted from the [embed] lever. .dockerignore is a licence boundary: an image layer is a copy and a pushed image is redistribution, so media and .env stay out. uv.lock is no longer ignored, because a lockfile that is not in the repo pins nothing.

Numbers, with the commands

uv run pytest tests/unit -q                                671 passed, 1 skipped   (was 645, 1)
make leakage-check                                         PASS, overlap 0
uv run pytest tests/gates/gate_phase0.py gate_phase1.py    17 passed
uv run pytest tests/gates/gate_phase2a.py -q -s            10 passed
make overview VIDEO=611 OVERVIEW_FLAGS=--refresh           6 windows, 3m16s, openai/gpt-oss-120b
  people 24  evidence not an exact chunk range: 0
  topics 72  start not a real chunk start, or end past 1804.8s: 0
docker compose config                                      exit 0

gate_phase2a was re-run specifically because answer.reasoning_effort is a global lever and that gate scores the extractive path. It did not move: schema-valid 1.0000 (15/15), abstentions 3/3, abstention rate 0.1667 on 12 answerable against a 0.25 ceiling, 10 citations all grounded, 0 repairs, $0.0000.

What a reviewer should push on

• The image is not built. Docker Desktop's daemon is not running on this machine, so make docker-build && make docker-up && make docker-doctor is unrun. That is the first thing to try.
• config.deploy.toml is missing. src/config.py reads one file and refuses defaults, so a deploy config is either a 900-line copy that drifts or a small overlay needing load(path, overlay=). Left undone rather than done badly; only needed for a public host, since local compose already binds 0.0.0.0 through the existing HOST override.
• Items 4, 7, 8, 9, 10 of docs/plan-whole-video-questions.md are still open — the index_video hook, mode on AskRequest/Provenance, the mention._label fallback, the frontend control, and evals/overview/ + tests/gates/gate_overview.py. There is no gate on the overview yet, so its quality is argued from one run, not measured.
• 72 topics for a 30-minute video is granular (~25s each). Defensible, since each window's topics are contiguous, but worth an opinion.