feat(nudge): persist nudge injections as display-only session entries + single-line acp.log (closes #326)

[ranxianglei]

Closes #326.

Triage / verification

• Valid feature, no duplicates. [bug] Post-compression sent-view undercounts prompt by ~70–80K → false idle → host-floor emergency re-nudge loop #325 is the related-but-distinct bug (post-compression sent-view undercount driving repeated emergency injections); Title: Context never auto-compresses when the model ignores the nudge — climbs 75%→95%+ on long agentic tasks (no enforcement path) #269/fix(nudge): same-turn growth re-inject after a full growth floor (closes #269) #316 cover nudge enforcement, not persistence.
• Verified against pi-coding-agent 0.83.0 source:
  • pi.sendMessage({customType}) mid-turn is unsafe — while streaming, agent-session falls back to steer() (injects as steering input). The /acp precedent (commit 0ff701e, issue Show /acp-status as persistent output instead of a notification #255) only works because slash commands run idle; it cannot be reused from the context handler.
  • pi.appendEntry(customType, data) appends {type:"custom", ...} to the session JSONL — pure sync append, survives restarts, never sent to the model.
  • entriesToCoreMessages projects only type:"custom_message" entries into the sent view ⇒ type:"custom" records are display-only by construction. This satisfies the issue's "prefer a display-only entry type" constraint, so no debug gating is needed (zero context cost).
  • TUI: unrendered custom entries are dropped by addCustomEntryToChat, and the restart-rebuild path (renderSessionEntries) includes them ⇒ registering an entry renderer makes records visible both live and after pi --session <id> resume.

Changes

• Persistence (src/index.ts): on nudge injection, append a compact one-line record (e.g. [ACP nudge] EMERGENCY 95% · T1 · top range m00120–m00168) via pi.appendEntry("acp-nudge", {text}). Deduped per user turn — emergency nudges re-inject on every LLM call until retry-capped and must not flood the session file. markNudgeRecorded runs only after a successful append (a failed append retries on the next fire). Recorded-state cleared at session start alongside nudge-shown tracking.
• Renderer (src/index.ts): registerEntryRenderer("acp-nudge", ...) renders the record as a dim boxed line (pi-tui Box/Text, customMessageBg background). @earendil-works/pi-tui added as devDependency pinned exactly 0.83.0 (matching the host) and bundled inline by tsup — dist stays zero-runtime-deps (verified: no bare pi-tui import remains in the bundle).
• Single-line log (src/log.ts): \r/\n in string fields (including Error.stack) are escaped to literals, so every acp.log entry occupies exactly one physical line — grep nudge-injected now returns the full nudge body instead of just the headline.
• Docs: README Logging section (nudge audit trail + one-entry-per-line guarantee), CHANGELOG.

Tests (4 new; full suite green)

• formatNudgeRecord unit — exact-string assertions matching the issue example (emergency/non-emergency, tier, missing ranges).
• Integration — emergency overflow recipe (180K window vs ~270K sent view): renderer registered at factory time; exactly one persisted record per user turn across three context fires (emergency re-injection still reaches the model but must not duplicate the record); a new user message yields the next record.
• Renderer unit — well-formed record → box with one text child; empty data → undefined (invisible but persisted).
• log.test.ts — multi-line field values (nudge text + stack trace) stay on one tagged physical line.

Verification

• npm run typecheck ✅
• npm test ✅ — 466 pass / 3 skipped / 0 fail (469 total)
• npm run build ✅ — dist 627 KB, pi-tui fully inlined
• Smoke-loaded the built dist/index.js in plain node ✅