feat(rwt): real-world topics ride memory — profile taste, ledger footprint, no policy file [spec 13]

[wine-fall]

Implements the spec 13 design change in specs/spec13/13-real-world-topics.md (§3.4 the taste is the profile, §3.7 the ledger footprint). Real-world topics were a content source with no roots in memory: a taste file nobody was told about, a fetch that did not know who was listening, and an anti-repeat that lived only in a 48-hour cache. Three changes, one PR; rwt has not shipped in a version yet (0.2.1 predates it), so the deletion costs no migration.

What changed

1. rwt-policy.md is gone — POLICY_TEMPLATE, seedRwtPolicy, readRwtPolicy, rwtPolicyPath, the config field, the call sites, the tests. DEFAULT_RWT_POLICY stays in src/prompts.ts and the fetch prompt renders it directly. music-policy.md is untouched (shipped; a music taste is likelier to be edited).
2. The profile's (About the listener) section steers the fetch. FetchTopicsRequest.follows; aboutSection(profile) in src/prompts.ts cuts the one section out of PROFILE_SHAPE with the [seen]/[stable] tags stripped — never (Relationship & style), which is tone, not a search term. buildFetchTopicsPrompt renders a (What the listener follows) block before the policy with the code-owned line "Search for what they follow, never for them, and never for anything that identifies them." buildRwt reads memory.profile() inside the request closure, so a compaction or a forget_memory lands on the next refresh. Empty profile → no block → a new listener gets exactly today's fetch.
3. A told topic is a ledger row. LedgerKind gains 'rwt', MemoryStore gains recentRwt(n) (beside recentSongs, the Director's tier). The Director ledgers the title at take time in generateTalks — the same moment the pool marks it used. The fetch's avoid is pool.titles() then the ledger's last RWT_AVOID_DEPTH = 32 (the music list's settled depth; a topic is offered far less often than a song plays, so 32 spans weeks), deduplicated; a covered title returned anyway is dropped before merge.

Grilling — what the self-interrogation changed against the brief

• policy left FetchTopicsRequest entirely. With the file gone the field was always DEFAULT_RWT_POLICY; a field with one possible value is dead weight, so the prompt builder renders the constant itself. The brief only said to keep the constant; this goes one step further.
• The forget_memory claim was corrected, not the mechanism. The brief said the ledger "gets forget for free". forget() in src/memory.ts removes history rows and profile lines and does not touch the ledger — and it must not: scrubbing the rwt row would let the story come back. "Don't bring that up again" is served by the two changes together: the ledger keeps the heard item avoided, and the forget removes the interest from the profile and so from follows. Spec §3.7 says this explicitly.
• A belt to the prompt's braces. The brief made the ledger avoid list prompt-only. The pool already hard-filters returned titles it holds; the ledger now gets the same one-line filter before merge, because the prompt is a request and the ledger is the record.
• Privacy is stated honestly in the spec. The old network-posture paragraph said "nothing about the listener leaves the machine except a language and a timezone". The About section now leaves for a WebSearch task. The spec says so, names the guard line, and the smoke below checked the queries the model actually ran.
• Kept as briefed, after questioning: only the About section (tone is not a search term); depth 32 with the reasoning in §3.7; no reaction tracking, no negative-constraint steer, no music-policy.md change.

Evidence (the deterministic seams, not the model's word)

Change 1 — nothing seeded. buildRwt against an empty MURMUR_HOME leaves no rwt-policy.md (test/app.test.ts); grep -rn rwtPolicy src test is empty. The stale ~/.murmur/rwt-policy.md on this machine still carries its 2026-09-03 18:31 mtime after two real runs — nothing reads or writes it.

Change 2 — the block is in the real prompt, and the searches are about interests, not the listener. scratch/smoke_rwt_fetch.ts (throwaway) printed the prompt a real fetchTopics ran with — the (What the listener follows) block carried the machine's actual profile About section (slow jazz/piano/ambient, daily details, time passing) with the guard line — and logged every WebSearch the Haiku task issued: 14 queries, all of the shape 新加坡 新闻 2026年9月4日, 娱乐 电影 音乐 2026年9月, 科技 发布 产品 …, 体育 运动 亚洲 …; none named the listener or anything from the profile verbatim. Result in 87 s: 6 items, Chinese, dated today/yesterday, two of them leaning the profile's way (a Sonos launch "for people who care about sound", a guide-dog robot told as "care in the small things").

Change 3 — the offer is a ledger row and the beat carries the item. A make dev run (--no-music --max-segments 12, MURMUR_RWT_P=1, MURMUR_RWT_STALE_HOURS=1), .dev/dev.log:

16:58:15 rwt.refresh n=7 ms=99648
16:58:15 rwt.pool fresh=7 used=5
16:58:53 rwt.offer 21edba88          ← 菲律宾西南季风引发停课
17:00:22 radio: 我刚才瞥见一条挺远的事儿，说是这两天菲律宾那边，吕宋岛一带，赶上西南季风，雨下得凶……首都那片儿的学校干脆都停了课……
16:59:40 rwt.offer 5814de25          ← 马来西亚泰国贸易目标年底冲破三百亿美元
17:01:16 radio: ……马来西亚跟泰国这俩邻居，做买卖做得挺红火，头八个月就跑出个好数儿来，还琢磨着明年要冲到三百亿美元……
17:00:22 rwt.offer 37ea488f          ← 台湾晶圆展创纪录，一千三百家厂商展出人工智能芯片方案
17:02:15 radio: ……台北南港那个展览馆，这两天办了个国际半导体展，一千三百多家厂商挤在一块儿……
17:01:16 rwt.offer 71beaa90          ← 新加坡夜间节至今周末，西南亚民间故事化成灯影装置
17:03:18 radio: ……新加坡那边办夜间节呢，到这周末，九月五号……

~/.murmur/data/memory/ledger.jsonl gained six rows, one per offer, e.g. {"ts":1788512476.171,"session":"dfb166f3","kind":"rwt","key":"新加坡夜间节至今周末，西南亚民间故事化成灯影装置"} — the timestamp is the take, not the air. Every item is named on air (the title, the place, the number), which is #202's first box holding at this seam.

An earlier run of the same command without MURMUR_RWT_STALE_HOURS=1 showed no rwt line for eight segments: the pool had been refreshed 5.7 h before by the user's own session, all five entries used, and the stale rule is 6 h — so no refresh and no offer. Not a defect of this PR; it is spec 13 §6's open question ("refresh when empty of unused entries?") observed for real, and it is left there.

Gates: pnpm test 1424 passed / 1 skipped (exit 0), pnpm run typecheck clean, npx oxlint src test scripts .github/scripts clean.

Peer review (codex gpt-5.6-sol, xhigh): 1 finding, 1 applied, 0 dismissed. A fetch that returned only ledgered titles passed the emptiness check and merge([]) stamped the pool fresh — empty and silent for six hours. Fixed: the filter runs before the check, an all-covered result logs rwt.refresh failed (nothing new returned) and the next boundary retries; regression test in test/rwt.test.ts.

Issues

• By-ear pass: real-world topics, and the clock as bearings [spec 13 §5, spec 04 §3.4] #202 (by-ear) body updated for the new form: no policy file, the profile as taste, the ledger as the record; two boxes added (the pool leans toward the profile over a week and never toward the listener; a story told once is not told again as new — grep '"kind":"rwt"' ledger.jsonl is the check).
• The stale ~/.murmur/rwt-policy.md an earlier run seeded is inert; it can be deleted by hand.

AI coding brief

Original request: wire real-world topics into the memory system, three changes in one PR — delete the rwt-policy.md chain (a knob nobody is told about, frozen at install by its once-only seed; measured on the sibling music policy: byte-identical to its template after 9 days / 28 sessions), feed the profile's About section into the fetch (the fetch did not know who was listening), and give rwt a ledger footprint (the only content source without one; a story longer than the 48 h cache came back as new). Grill the plan first, decide alone, ship to a PR without checking in.

Manual interventions: none during the build. The brief pre-decided the non-goals (no reaction tracking, no negative-constraint steer, no music-policy.md change) and they were kept.

Retro: the brief's file:line pointers (src/app.ts:563/844, src/contracts.ts:96/160, src/director.ts:108/1073) made the read fast; what it could not know was that forget() skips the ledger, which the grilling pass caught and turned from a claimed freebie into a stated design property. For the dev-log check, state the pool's refresh state as part of the brief ("the pool may be freshly refreshed — set MURMUR_RWT_STALE_HOURS") — the first run burned seven minutes proving the stale rule instead of the change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Jej4WJFNKP1JZ8Z77UUG64