Merge #48: restore Omni frame/first+last/reference on the batch API

[crisng95]

Lands #48 on top of the REST removal (#49), with its REST half removed rather than revived.

Why this branch exists instead of a rebase

#48 was cut from a49831f, before #49 removed the REST transport, and it preserved the dual-transport shape deliberately — "the legacy REST implementations are kept below for non-batch deployments." On current main that half does not merely conflict, it does not import:

ImportError: cannot import name 'USE_BATCH_RPC' from 'agent.config'

agent.config.USE_BATCH_RPC, agent/services/headers.py (random_headers) and FlowClient._build_url/_client_context are all gone. 12 conflict hunks across omni_flash.py (4), test_omni_flash.py (7) and docs/OMNI_FLASH.md (1).

What the resolution keeps

#48's batch submits, byte for byte — the captured envelopes eb1hJf (first frame), nprQif (first+last), MZZa6b (references), and the 360p model-key suffixes. A merge resolution should not quietly reshape a contributor's verified wire payloads.

What it deletes

• the REST bodies of _submit_omni_frame_video and generate_omni_flash_video, the tRPC poller behind check_omni_flash_status, plus _annotate_polling, _fetch_project_initial_data, _fetch_media_url
• the two surviving if USE_BATCH_RPC: blocks (unwrapped), and the flag's 5 imports / 4 references
• both model_key = _load_model_key(...) assignments — their only consumer was the REST body's videoModelKey. _load_model_key itself stays (tested directly), and models.json still carries all three modes × 4 durations, so no real validation is lost

omni_flash.py goes 709 → 417 lines; imports return to main's set.

Tests — two needed judgement, not deletion

#48's suite was pinned to REST by a module-level autouse fixture setting USE_BATCH_RPC=False, so tests that never name the flag were still exercising the old poller. My first pass misclassified three of them and the test run caught it.

• dropped 7 REST/tRPC wire-contract tests
• converted ..._returns_signed_url_without_buffering_video and ..._missing_media_is_pending onto the batch poller — what they pin still matters: the poller must hand back Flow's signed url instead of buffering the clip, and a media record without a /video/ url is pending, not done
• dropped ..._requires_project_id_for_legacy_descriptors — the batch poller deliberately resolves a missing project id instead of refusing
• rewrote the seven-reference boundary: main asserted it reached the capability gap, Restore Omni first/last/reference modes on migrated Flow API #48 asserted it through the REST body. Neither applies now, so it asserts the batch wire. Seven is the limit, not one past it

272 passed (up from 266), suite exits cleanly, agent.main imports.

Docs this merge falsifies

The merge invalidates claims main shipped in #49, so they move with it: README's top note and capability table, CLAUDE.md, and the UNSUPPORTED_ON_BATCH_API: Omni Flash row in /fk-doctor — an error string nothing can produce any more.

The unported set is now three, all on the Veo path (flow_client.py is untouched by this merge): video upscale, Veo r2v, Veo start+end chaining. Callers hitting the Veo guard are now pointed at the Omni modes that cover the same shot. CLAUDE.md's "4K upscale" was corrected to video upscale — it read as the image export, which works.

Not verified

The captured envelopes themselves. #48 reports them live-verified on 2026-09-14 and the tests lock their shape, but proving the wire format needs a real submit through a signed-in Flow tab, which this branch did not do.

Merging this closes #48 — the merge commit carries b1db615 as its second parent.

🤖 Generated with Claude Code