fix(persona): always-record persona turns including failures

[joelteply]

Summary

Per RTX (continuum-8e97)'s #1068 review finding: `record_turn()` at response.rs:273 only fired on the `Ok`-path of `respond()`. If `respond()` returned `Err` (analyze fail, render fail, any internal `?` propagation), NO fixture landed on disk — direct contradiction of recorder.rs's F1-telemetry header ("F1 cars don't ship without telemetry") and Joel's "evidence is for the debugger, not the trash" rule. The most diagnostic turns (the failures) left zero trail.

What changed

• `respond()` wraps `respond_inner()` with always-record. Inner-fn pattern carries every `?` exit point so the outer fn captures Ok with the real `PersonaResponse` OR Err with a synthesized `TurnError` recorder artifact + partial `CognitionTrace`.
• `recorder.rs` adds `record_failed_turn(input, error_msg, total_ms, trace)` writing `{rustResponse: null, rustError: TurnError{error_msg, last_completed_seam, partial_trace_seams, total_ms}, cognitionTrace}`.
• `PersonaResponse` enum stays pure {Silent, Spoke} — no TerminatedAtSeam widening. Recorder-only artifact preserves separation: runtime value type for success-path semantics; debug fixture for diagnostic.
• `CognitionTrace` adds `last_seam_name()` + `seam_count()` helpers consumed by the failure-path synthesis.
• TS `PromptCapture.ts` sink + `SentinelCleanupServerCommand.ts` tombstone deleted — Rust recorder is now SSoT for capture (Joel: "stdout/println invalid for realtime systems").
• 3 integration test files get `other_persona_names: Vec::new()` to fix pre-existing canary compile drift in test fixtures (PR Persona resource substrate + native multimodal restoration #950 added the field, fixtures didn't update).

Validation class

• Contract TDD: yes — `record_failed_turn_writes_error_with_partial_trace` asserts fixture shape, `last_completed_seam` correctness, partial trace survival.
• Failure TDD: yes — same test simulates "render adapter timed out at 30s" with only SEAM_ANALYZE recorded; asserts that seam survives in the fixture.
• Performance/Resource/Residency VDD: n/a — observability-only diff, no hot-path changes.
• Browser/UX evidence: precommit browser ping passed.
• Platform coverage: Mac M5 Pro Metal, cargo test --lib --features metal,accelerate.

Validation

• `cargo test --lib --features metal,accelerate` — 1922 passed; 0 failed; 28 ignored (includes new `last_seam_name_tracks_most_recent_record`, `seam_count_matches_recorded_seams`, `record_failed_turn_writes_error_with_partial_trace`).
• Precommit hook: TypeScript clean; browser-ping precommit test passed.
• Pre-push hook: Rust lib tests passed (64s); Docker native push skipped (worktree had unstaged test fixes at the time — re-runnable per scripts/push-current-arch.sh).

Three-way convergence

• @airc-8a5e (Mac sibling) flagged scope match on inner-fn + always-record pattern; Q1 architectural ACK.
• @continuum-8e97 (Windows/RTX) original Make Rust persona recorder the single fixture source #1068 reviewer; ACK with principles review (recorder-only failure capture = correct SSoT, no PersonaResponse widening, partial trace as diagnostic value).
• @airc-8a5e (Codex) implementation execution + final code review.

Known gaps

• Pre-existing canary breakage in `tests/qwen35_live_pipeline_diff.rs` (SamplingConfig API drift) and `tests/persona_prompt_token_diagnostic.rs` (format-string syntax) is NOT addressed here — separate PR scope.
• Followup work: SEAM_RAG_COMPOSE addition + RagSourceTrace{source_id, snippet_hash, score, included|dropped, drop_reason, bytes_in_prompt} in trace.metadata, raised by the doc-vs-shipped-Rust review of docs: define CBAR-like Rust substrate #1081 (CBAR substrate).