Move the eval bench's direction probe onto per-line direction and report stripped events honestly

[schmug]

Task

Update the TTS eval bench (skills/tts-eval/bench.py) for what #201 changed in the renderer: its direction probe should exercise per-line direction over a clone (a lines scene whose one line carries instruct), not only the episode-level voice_instruct path, and its events probe should report that an engine without events never received the markers (they are stripped before the render) instead of a bare markers heard: none that reads as "performed".

Context

#201 made render.py consume the events and direction capabilities: on an engine without events the renderer strips (laugh) / (sigh) from every take before the model sees it, and a scene line may carry instruct, rendered through Breeze's clone-plus-instruct form and folded into the take's cache key. The bench renders only through render.validate_manifest / render.render_segments (the invariant in CLAUDE.md), so two of its probes now measure something different from what they were written for:

• The events probe on the control (qwen3) measures a marker-stripped take: WER 0, markers_heard: []. tests/test_tts_eval.py::test_bench_renders_candidate_and_control_and_writes_both_artifacts already asserts that per engine, but the ledger row and the HTML report present it the same way as a candidate that performed the markers. A reader comparing engines cannot tell "performed" from "never handed them".
• The direction probe goes through voice_instruct (design mode on a plain segment), which skills/tts-eval/SKILL.md line 28 explicitly says is a stand-in until Gate vocal events and per-line voice direction on the engine's capabilities #201 lands. The identity drift that motivates st_write.MAX_DIRECTED_LINES_PER_SCENE = 1 (Ethan 93 → 183 Hz, similarity 0.96 → 0.88 in the 2026-09-04 eval) needs to be measured on the form a show actually uses: a cast clone with instruct, compared against the same clip undirected.

Pointers

• skills/tts-eval/bench.py — TAKE_KINDS, strip_events, the events/direction entries in the corpus expansion (~line 195-235), markers_heard in the per-take rows (~line 674-723) and the report (~line 951, ~line 1005).
• skills/tts-eval/corpus.json — the events and direction entries (append-only; a changed line makes earlier ledger entries incomparable).
• skills/tts-eval/SKILL.md — the probe table and the metrics table, both pinned to bench.METRICS / the corpus by tests/test_tts_eval.py drift tests.
• skills/daily-podcast/render.py — EVENT_MARKERS, strip_event_markers, _validate_engine_capabilities (the per-line instruct refusal), _generate_breeze (the clone-plus-instruct form), _segment_cache_key(instruct=).
• skills/surface-tension/st_write.py — DIRECTIONS (word → instruct phrase) and lines_for_engine; the probe should direct with a phrase from that table so the bench measures what the show would ship.
• tests/test_tts_eval.py — the per-engine events assertion added in Gate vocal events and per-line voice direction on the engine's capabilities #201's PR, and the fake engine/analyzers the bench tests run on.

Constraints

• The bench keeps rendering through the registry only (test_bench_never_calls_mlx_audio_directly stays green).
• The corpus stays append-only and under every finite max_take_chars in render.ENGINES; adding a directed variant of an existing line is fine, changing a line is not.
• On an engine without direction the per-line probe is skipped and the report says so (the existing posture for an engine without design).
• Ledger entries written before this change must still parse; a new field is added, never renamed.

Acceptance criteria

• The events probe row and report cell distinguish "performed" (engine has events, markers absent from the transcript) from "stripped before the render" (engine lacks events), and the daily-skill metric name for it is pinned by the existing drift test.
• The direction probe renders a clone line with instruct (a DIRECTIONS phrase) through render.render_segments and reports its pitch/similarity against the same line undirected.
• skills/tts-eval/SKILL.md no longer says the direction probe waits on Gate vocal events and per-line voice direction on the engine's capabilities #201.
• pytest and both ruff gates pass.

Out of scope

• Changing what render.py strips or refuses; that contract is Gate vocal events and per-line voice direction on the engine's capabilities #201's.
• Adding new markers or direction words to the show's closed lists.
• Switching Surface Tension to Breeze (Switch Surface Tension to Breeze-TTS-2 after the ear test #203).

Cross-references

• Gate vocal events and per-line voice direction on the engine's capabilities #201 (events and direction, gated on capabilities)
• Add a tts-eval bench skill over the engine registry #200 (the bench)
• feat(tts-eval): bench a TTS engine from the registry against the production engine #209 (the bench landing)
• Switch Surface Tension to Breeze-TTS-2 after the ear test #203 (the Surface Tension engine switch)

Generated by Claude Code.