t27f: R4f -- mode 5 holds the address for REEL_CTA_MIN_MS on a sped-up voice #10629

Follow-up to R4d (#10614, PR #10627), epic #10560.

In LR_MODE=5, crossfade 1 is centred on cta_from, the moment the address takes over. The address is now kept off the crossfade, so it only comes in after the crossfade ends. On a voice the reel had to speed up, that leaves it on screen for less than REEL_CTA_MIN_MS. A test in PR #10627 measures this cost: it is real, and it is at least MIN_SCENE_FRAMES - 1 frames.

Boundary

• specs/t27f/lesson_reel.t27 and its seal: the beat cut for the address, or the address window, in mode 5.
• Out of scope: modes 3 and 4. If the fix moves crossfade 1 in mode 4 too, that is a decision to name in the PR with its own cmp evidence, not a side effect.

User Scenarios

• A viewer of a mode 5 reel narrated by a fast voice still has at least REEL_CTA_MIN_MS to read the address before the reel ends.

Requirements

• Mode 5 holds the address for at least REEL_CTA_MIN_MS on every voice the clock accepts, and it is still never on a crossfade.
• The candidate fix is to end crossfade 1 at cta_from instead of centring it there. If that is the fix, the change to mode 4's cut is stated and tested.
• The test "the cost: on a voice the reel speeds up, mode 5 holds the address for less than REEL_CTA_MIN_MS" is replaced by a test that the hold holds, plus a control showing the old cut fails it.

Success Criteria

• On the Railway lab, t27c test-report passes with 0 vacuous tests, and the seal is saved and verified.
• A grid test over every lesson, both languages and every voice the clock accepts shows the mode 5 address hold is at least REEL_CTA_MIN_MS and is never on a crossfade.
• The cmp regression for mode 3 finds 0 differences. For mode 4, either 0 differences or a counted, explained diff that is limited to crossfade 1.

Requirements

FR-001: Mode 5 MUST hold the address for at least REEL_CTA_MIN_MS on every voice the clock accepts.
FR-002: The address in mode 5 MUST never appear on a crossfade.
FR-003: If the fix ends crossfade 1 at cta_from instead of centring it there, the change to mode 4's cut MUST be stated and tested.
FR-004: The existing test that measures the cost (mode 5 holds address for less than REEL_CTA_MIN_MS on a sped-up voice) MUST be replaced by a test that the hold meets REEL_CTA_MIN_MS, plus a control demonstrating the old cut fails it.

Success Criteria

• t27c test-report on the Railway lab exits 0 with 0 vacuous tests, and the seal for specs/t27f/lesson_reel.t27 is saved and verified.
• A grid test over every lesson, both languages, and every voice the clock accepts shows mode 5 address hold ≥ REEL_CTA_MIN_MS and the address is never on a crossfade (measured in frames).
• cmp regression for mode 3 shows 0 differences. For mode 4, either 0 differences or a counted, explained diff limited to crossfade 1.