[ffroliva]

Summary

• _select_video_duration now raises UiSelectorDriftError (typed, exit code 23 — the gflow image t2i fails finding Flow aspect-ratio mode switch trigger #183 selector-drift pattern) with a debug screenshot (debug_no_duration_tab.png) instead of warn-and-continue when the duration_tab probe misses.
• No behavior change when --duration is omitted: the function only runs for an explicit duration (classic.py guards on request.duration is not None), so Flow's default path is untouched.
• No new error class or exit code — reuses the existing UiSelectorDriftError that sibling probes (_switch_video_sub_mode) already raise for the same failure mode.
• Error detail carries a remediation hint (omit --duration to accept Flow's default) and the screenshot path, which doubles as the investigation artifact for the selector-drift question below.

Why

Duration is a contract parameter: a silent 4s→8s substitution corrupts downstream timeline math while the JSON result reports success (issue #288, 3/3 repro on the 2026-07-11 chalkboard i2v pilot). Same design principle as the #281 wrong-media fixes: an unhonorable explicit parameter is a hard error, not a warning.

Verification status

• TDD: failing test written first (test_missing_duration_tab_raises_drift_error), then the fix to green
• Unit: raise-on-miss, screenshot capture + Screenshot: clause, remediation hint in message
• ruff check + ruff format --check clean repo-wide; pyright 0 errors; 1271 passed / 3 skipped (tests/api + tests/cli)
• Needs human e2e: live i2v run with explicit --duration on a Classic profile to confirm the happy path still sets the tab, and to capture the new debug screenshot for the open selector-drift investigation (why duration_tab misses on the current Frames-submode DOM — possible cohort split, cf. gflow image t2i fails finding Flow aspect-ratio mode switch trigger #183). Headed browser + Veo credits required.

Scope notes

• The selector itself is unchanged — investigating/repairing the duration_tab cascade against Flow's current DOM stays open on video i2v: --duration is silently ignored when the duration tab probe fails (requested 4s, got 8s) #288 (needs the live DOM; the screenshot this error captures is the input to that work).
• --best-effort opt-out flag skipped (YAGNI until someone asks); requested_duration/effective_duration payload fields unnecessary under fail-fast.

Refs #288

🤖 Generated with Claude Code

[ffroliva]

Council review verdict (6 dimensions @ 868e8c8): YELLOW → findings applied in 02dcf01

Dim	Verdict	Headline
D0 CI-mechanical	🟢	ruff check/format, doc links, hygiene all green locally
D1 Correctness	🟢	Single choke point; nothing swallows the error (MCP → structured error, worker → task fail); no warn-path remains that drops duration
D2 Quality	🟢	Matches sibling drift-raise pattern; 2 cleanups flagged
D3 Security	🟢	No new surface; PII screenshot warning + show_locals=False renderer inherited
D4 Tests	🟡	Classic-driver call site (out_dir forwarding, propagation) had zero coverage
D5 Memory/docs	🟡	CHANGELOG + KNOWN_ISSUES entries missing per #281-wave convention
D6 UI/live-verify	🟡	Affected surface unverified (declared); locale-label hypothesis for the selector rot

Applied in 02dcf01: new tests/api/transports/drivers/test_classic.py (D4 must-fix); CHANGELOG [Unreleased] + KNOWN_ISSUES Open entry with the locale hypothesis (D5 must-fix); dropped the double-signaling pre-raise duration_not_set warning, dropped the out_dir default, fixed the misleading aria-label docstring, message now says 'typically 8s' (D1/D2/D6 nice-to-haves). Locale hypothesis also posted to #288.

Deferred (follow-up): wrap UiSelectorDriftError into ChainPartialError in chain.py:195 so a mid-chain drift abort reports completed links (D1, pre-existing for sibling raises).

Remaining YELLOW (intrinsic): the live-surface box — one headed Classic-profile run with explicit --duration (happy path + the real miss screenshot). That is the human-e2e gate already declared in the PR body; keep draft until checked. SonarCloud gate pending on CI at time of review.

[ffroliva]

Live e2e result (2026-07-11, denon82, credit-free) — fail-fast verified on the affected surface ✅

gflow video i2v <local file> "…" --duration 4 (correlation b8e66830):

• duration_tab probe missed (reproducing the pilot's 3/3), and the new code aborted pre-submit with exit 23 — the settings popover showed 'A geração vai usar 10 créditos', so the fail-fast saved that spend instead of producing a silent 8s clip.
• debug_no_duration_tab.png was captured and answers the video i2v: --duration is silently ignored when the duration tab probe fails (requested 4s, got 8s) #288 investigation: the duration control is absent from this cohort's Frames-submode popover entirely (mode/sub-mode/aspect/count/model rows only). The locale hypothesis is refuted — the UI is Portuguese and the sibling count tabs match fine. Details + open question (removed vs moved) posted on video i2v: --duration is silently ignored when the duration tab probe fails (requested 4s, got 8s) #288; KNOWN_ISSUES corrected in b3552fd.

Happy-path caveat: the 'duration tab found → clicked' path is unverifiable on this profile — the control doesn't exist on this cohort. If another profile/cohort still renders duration tabs, one credited run there would close it; otherwise the fail-fast IS the permanent correct behavior for affected cohorts.

The PR's verification box can be considered satisfied for the reachable surface; evidence log in the session scratchpad (e2e289.log).