feat: AI-companion reply-quality lightweight layer (Spec 2, issue #84)

[enriquephl]

Summary

Six small, independent prompt/quality changes that make companion replies less empty / templated / self-absorbed / out-of-character — no new architecture, no schema migration, no API surface change. Implements docs/superpowers/specs/2026-06-13-companion-reply-quality-design.md (Spec 2 of the issue #84 follow-up).

1. Sampling params — top_p / frequency_penalty / presence_penalty (Option<f32>, task-level) plumbed config → wire (sync + stream) → pipeline, with opinionated defaults in examples/model_config.toml (0.9 / 0.4 / 0.2). None ⇒ omitted from the wire, so a deployment setting none produces a byte-identical OpenRouter body. No repetition_penalty (provider-inconsistent, distorts CJK).
2. Dynamic anti-repetition — new pure repetition::overused_openings mines over-used reply openings from the persona's recent assistant turns (ChatRepo::recent_assistant_contents), injected as a volatile [avoid_repetition] block.
3. Anti-templating iron rules — directives ⑨/⑩/⑪ against self-narrated gaze openings (我看着…), ellipsis overuse, and consecutive 我-openings.
4. Memory-extraction specificity — rewritten memory_extraction / insight_extraction filter_prompts push for concrete, evidenced memories (good-vs-bad examples); 5-category + anti-attribution + JSON contracts preserved.
5. Persona guard clause — a constant PERSONA_GUARD re-appended after the persona's authored system_prompt (no AI-acknowledgment / no meta-leakage / no self-censorship), anchored "within all other hard constraints" so it cannot override iron-rule ⑦ (minor-safety).
6. Emotional context — last 5 affinity-evaluation reasons (AffinityRepo::recent_emotional_reasons) rendered oldest→newest as a volatile [emotional_context] block.

Invariants preserved

• Cache prefix: PERSONA_GUARD is a constant in the stable prefix; the two new sections are volatile (after the [turn_style] cut). Empty inputs render byte-identically to before.
• Non-fatal: both new store fetches degrade to empty on DB error (warn-level), mirroring recent_turn_pairs.
• Deferred (NOT built): reply-action decision layer; memory salience/evidence columns.

Test Plan

• cargo fmt --all -- --check clean
• cargo clippy --workspace --all-targets -- -D warnings clean
• cargo test --workspace --all-features — 554 tests, 0 failures (incl. new unit tests for the miner, both store fetches, the two prompt sections, the iron-rule directives, PERSONA_GUARD, sampling resolve/wire, and cache-prefix invariants)
• OpenAPI snapshot — no drift
• codex review

🤖 Generated with Claude Code

[enriquephl]

Codex review (codex exec review --base dev) — response

Two P2 findings on the emotional-context feature (Component 6):

• P2 — recent_emotional_reasons lacked a turn-boundary cutoff → fixed in a047c8f. Added a before_message_id cutoff (e.created_at < (current user row's sent_at)), mirroring ChatRepo::recent_assistant_contents and the sibling recent-chat queries, so a racing concurrent same-session stream can't leak a later turn's reason into an earlier prompt. New test recent_emotional_reasons_excludes_events_after_cutoff covers it.

• P2 — preceding turn's reason may not be persisted before the next message → accepted as by-design. The affinity event is written by the detached post-process task (≤10s evaluator timeout), so a rapid next message reads a trajectory that lags by one turn. Codex's suggested fix (block/coordinate with post-processing) would couple chat-response latency to the affinity evaluator — disproportionate for a soft [emotional_context] hint that's explicitly "仅供参考，别照搬". The lag degrades gracefully and is inherent to reading async-written data; the P2 cutoff above doesn't worsen it.

Gate after the fix: fmt clean · clippy -D warnings clean · affinity tests green.