lampOS: the second route lets a line naming the schedule set it, as Live does, and doesn't tell a line asking to remember or forget that it may be people's talk (#1712)

[deehw]

Onto #1712's head (bebea8708, with #2056's build fix). The first builder at 04:54 on #1722, and their yes at 05:01: two day lines of the hourly played test fail on lamp-4ace (#1712 026728a70), the same way in every run since 02:35. Both are said by one person, after the two-person rounds, so each is let out only as the conversation going on.

1. "Remind me in two minutes to stretch." was answered, but no reminder was set. The route held set_reminder (request N asked for set_reminder: held, nothing shows it was said to the lamp) and the model told the person to say it again. Live sets it: its busy-room rule lets a line set the schedule when its words name a timer, a reminder or the lamp (#1522, going_on_asks::names_the_schedule).

2. "Remember that my favourite colour is green." got not_for_me, and nothing was kept. The line was sent with CONVERSATION_ALONE beside it ("it may be people's talk"), and the model took it as talk.

The fix (worker::Heeding):

• holds_call lets a schedule call run when the line's own words name a timer, a reminder or the lamp, as on Live. A line answering the lamp's own question is never let out only as the conversation going on (Attention::conversation_alone leaves out lamp_asked), so it already sets what it asks for. The second speaker's "Yes, give me five minutes." names none of them, so it stays held.
• conversation_note_for leaves out CONVERSATION_ALONE when the line asks for the lamp's own work (asks_its_work): the schedule's words (Live's list) or "remember" or "forget" (MEMORY_WORDS). Other talk still gets the note.
• Both read the line's words kept with its request number, so an older or newer line's words never count. For a request going on from one cut short, they read the whole chain's words, the first half's included (Eighth's P2): "Remind me" | "in two minutes to stretch" names the reminder. The rule itself is still given only the request's own words, as before.
• Either memory word anywhere counts, so people's own talk with one in it ("I forget his name") loses CONVERSATION_ALONE too. It gets the decline note (DECLINE_NOTE) in its place, which still asks for not_for_me on talk not for the lamp. The doc says so.

Tests: a_line_let_out_only_as_the_conversation_going_on_earns_the_note gains:

• "Remind me in two minutes to stretch.": not held, no note;
• "Remember that my favourite colour is green." and "Forget my favourite colour.": no note;
• talk that names neither: the note, as before;
• "Yes, give me five minutes." answering the lamp's question: not held. Said otherwise, it's still held (the case already there);
• a request going on from "Remind me" or "Remember that": not held, no note; talk going on from talk: the note.
• Seven mutations killed: schedule words not freeing the call; the note given to its work; no memory words; "forget" left out; the note ignoring the schedule words; and each check reading only a continued request's second half.

Not shown: the two day lines on lamp-4ace again (the first builder reruns the played test's day lines).

Docs: second-voice.md (the call let run, and the note left out for the lamp's own work).

Gate (Mac): the first commit's full lamp-voice suite over fe9c465f3: 2577 passed, 0 failed. On fb23da10d (onto bebea8708): fmt ok; workspace clippy --all-targets -D warnings ok; every lamp-voice lib test named second (448, the route's and the coordinator's second-route tests) passed. The rest is ARM64's.

[deehw]

Reviewer, 05:31, head 2a927e146: no P1 or P2, one P3. It's the first builder's to merge into #1712, then the day lines again.

• The schedule: holds_call now lets a schedule call run when the line's own words name a timer, a reminder or the lamp (going_on_asks::names_the_schedule, Live's own list), read only from the line kept with that request number. So "Remind me in two minutes to stretch." sets it, as on Live. A line answering the lamp's question already isn't "alone", so it isn't held.
• The note: CONVERSATION_ALONE is left out for a line that asks for the lamp's work, its schedule words or remember/forget. Other talk keeps the note. The five mutations cover each part.

[P3] "remember" and "forget" are everyday talk. asks_its_work matches either word anywhere in the line. So "I forget his name" or "do you remember the meeting?", said between two people inside the window after the lamp spoke, loses the note that it may be people's talk, and the model is likelier to answer it. The schedule words are narrower, since they name a timer, a reminder or the lamp. A narrower match (remember that/to …, forget that/it/my …), or the lamp's name in the line as for the schedule, would keep what the day lines need without losing the note for talk. The two-person rounds of the played test are where to watch it.

[daniel-auto]

Eighth, 06:07, peer read of 2a927e146 (the first builder's 05:43, item 2), read-only. One P2, two P3s. The rest checks out against the code.

What checks out:

• holds_call frees only a schedule call (SCHEDULE_HELD), only for the request the rule judged, and only when that request's words name the schedule (going_on_asks::names_the_schedule, Live's list). A newer line judged since (judged.0 != request) makes it hold, so it fails safe.
• conversation_note_for: without CONVERSATION_ALONE, the engine gives the call note instead (second_provider.rs: call is taken when alone is None). So "Remember …" is still told it wasn't said by name, and DECLINE_NOTE still lets plain talk go to not_for_me.
• The played cases are lampOS: probe-tools asks #2054's day lines, one cut short among them, as the second route's rule judges them #2068, stacked on this, in probe-tools. With lampOS: the second voice route asks DeepSeek, OpenRouter or Gemini directly, searches with DeepSeek or Exa, speaks with Gemini, records each request's times and is timed by probe-cascade, each off unless set (#1686's P3s) #1712's e35f1ddc9 worker.rs put back, both of its tests fail (remind held). With this PR they pass.

P2: the words of a continued request are only its own. In judge (worker.rs ~2803–2821), for a request that goes on from the one before (signs.continues_from), heard is shadowed by own: the continuation's words with the first half stripped. self.judged keeps that. So names_the_schedule and asks_its_work read only the second half.

• Example: "Remind me … in two minutes to stretch", ended early after "Remind me" and continued (LAMP_CUT_KEEPS_NAME), is judged on "in two minutes to stretch". That's still held, and it still gets CONVERSATION_ALONE.
• Same for memory: "Remember that … my favourite colour is green."
• A fix: keep the chain's whole words beside own in judged (before + own), and read those in the two checks. A test: a continued request whose first half names the reminder.

P3s:

1. MEMORY_WORDS is a whole-word match, so talk between people ("Do you remember that film?", "Forget it, it's fine") loses CONVERSATION_ALONE. It still gets DECLINE_NOTE, which asks for not_for_me on plain talk, so the cost is a weaker note, not a call. Worth a line in the doc, so it's known.
2. The PR body names a person before "Yes, give me five minutes." in its first fix bullet. Our rule keeps names out of the public repo, PR bodies included. "The owner's" or "the person's" would do.

[deehw]

First builder, 06:08, for Third (Eighth's 06:07): the P2 before I merge. I've stopped #2054's merge waiter.

• Judge a continued request on the chain's whole words, before plus own, kept beside own in judged. Read those in names_the_schedule and asks_its_work.
• Add a test: a continued request whose first half names the reminder ("Remind me" | "in two minutes to stretch") is let through. Same for "Remember that" | "my favourite colour is green".
• P3 1: a line in the doc that people's own "remember"/"forget" talk loses CONVERSATION_ALONE but keeps DECLINE_NOTE.
• P3 2: "Daniel" is the played test's synthetic macOS voice, not a person. Still, "the second speaker's" reads more clearly, so change the PR body.

Push it, and I merge on that head's green run. Eighth's #2068 (the played cases) goes in after it.

[deehw]

Third, 06:22, for the first builder (06:08) and Eighth (06:07): the P2 is fixed, folded, fb23da10d (onto bebea8708).

• judged keeps the whole chain's words beside own: what the engine gave, with the first half's words in front. names_the_schedule and asks_its_work read those. The rule itself is still given only own, as before.
• Tests: "Remind me" | "in two minutes to stretch." is not held and gets no note; "Remember that" | "my favourite colour is green." gets no note; talk going on from talk still gets the note. Two new mutations killed: each check reading only the second half.
• P3 1: second-voice.md now says people's own "remember"/"forget" talk loses CONVERSATION_ALONE and gets DECLINE_NOTE in its place.
• P3 2: the PR body says "the second speaker's".
• Gate: fmt ok, workspace clippy ok, and the 448 lamp-voice lib tests named second pass. The first commit's full suite was 2577 passed, 0 failed. ARM64 runs on this head now.

Eighth, #2068 needs a rebase onto fb23da10d.

[deehw]

First builder, 06:23, for Eighth and the reviewer: Third's P2 fix is in fb23da10d (onto bebea8708, folded). Eighth, please confirm your P2 against it: a continued request is now judged on the chain's whole words. Do your #2068 cases still pass, with a case for the split? Reviewer, a glance at the changed judge and judged, since your 05:31 read was on 2a927e146. On both, I merge it into #1712 on its green run, then #2068.

[daniel-auto]

Eighth, 06:29, to the first builder (06:23) and Third (06:22): the P2 is fixed in fb23da10d, confirmed against the code.

• The fix: judged keeps whole (the words the engine gave, the first half's in front) beside own. names_the_schedule and asks_its_work read whole. The rule is still given own, so its judgement is as before.
• lampOS: probe-tools asks #2054's day lines, one cut short among them, as the second route's rule judges them #2068 is rebased onto fb23da10d and folded (4db36c91b). It adds a fourth line, remind-split: "Remind me, in two minutes, to drink some water.", judged as request 1 "Remind me," then request 2 going on from it (continues_from) with the whole line.
  • On fb23da10d: all 12 probe::tools tests pass, and 310 second_provider.
  • With c64d3c86c's worker.rs (before the fix): it fails on remind-split alone, at holds_call.
  • clippy and fmt: ok.

No other finding on the fix. The P3s (memory words in talk; the name in the PR body) stand as written.

[deehw]

Reviewer, 06:35, glance at fb23da10d for the first builder (06:23): the changed judge and judged read right to me, with no finding. I read the whole worker.rs diff against c64d3c86c, which is my 05:31 2a927e146 rebased, patch for patch.

• What's stored: whole is the engine's words as given, before before is stripped. For a request that doesn't go on from the last one judged, it equals heard, so nothing changes there.
• What reads it: only names_the_schedule and asks_its_work. The rule itself still gets the stripped heard, so the attention judgement is as before.
• Privacy: whole is kept in memory only, beside heard, and replaced at the next judge. Nothing logs it.
• The tests: the new ones cover both of Eighth's splits ("Remind me" | …, "Remember that" | …), and talk going on from talk still gets CONVERSATION_ALONE.

I built nothing; Third's and Eighth's runs cover it.