lampOS: probe-tools asks #2054's day lines, one cut short among them, as the second route's rule judges them

[daniel-auto]

Stacked on #2054 (Third's fb23da10d, with the P2 fix): merge after it. The first builder's 05:43 on #1722, item 1: played cases for #2054, on the second route, each failing on #1712's head before #2054 and passing with it.

probe-tools gains the heeding's four lines (HEEDING_LINES). Each runs as the engine runs it:

• The rule: each is judged by the route's own rule (worker::Heeding, with --spoken-to), going on from a conversation in which the lamp asked nothing. So each is let out only as the conversation going on, as on lamp-4ace.
• The note: it's asked with the note the engine gives it then: conversation_note_for, else the call note, beside the line.
• The hold: its first call is put to holds_call.

The four lines:

• remind: "Remind me in two minutes to stretch." Right is set_reminder, not held.
• remind-split: "Remind me, in two minutes, to drink some water.", cut short after "Remind me," and gone on with as the next request (continues_from). Right is the same; it's judged on the whole chain's words. This was my P2 on 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) #2054, fixed in fb23da10d.
• to-someone: "Yes, give me five minutes.", said to someone else. Right is no schedule call run: none called, or held.
• remember: "Remember that my favourite colour is green." Right is remember, never not_for_me.

Functions: the model gets every function a spoken turn declares, plus the schedule's. The schedule's are read from the actions part's catalog, as a voice with --schedule-tools reads them. Where the catalog has no schedule, the lines aren't asked, and the probe says so; the other sections still run. It prints right=N/M and the functions called first per line: counts and names only.

Before and after #2054 (the_heeding_lines_are_judged_by_the_routes_own_rule, and the fake-chat test):

• With 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) #2054's worker.rs: both pass.
• 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 fail. remind's set_reminder is held.
• With 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) #2054's c64d3c86c worker.rs (before the P2 fix): both fail on remind-split alone.
• What's deterministic: the rule's part (remind held, remember told CONVERSATION_ALONE) fails on the head every time.
• What isn't: the model's not_for_me for remember under that note is DeepSeek's choice. The probe on lamp-4ace counts it.
• to-someone stays right on both, held. It's the guard that 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) #2054 doesn't let talk set the schedule.

Docs: second-voice.md, under probe-tools.

Gate (Linux):

• cargo test -p lamp-voice --lib -- second_provider: 310 passed, 0 failed (probe::tools 12);
• clippy -p lamp-voice --all-targets -D warnings: ok;
• fmt --check: ok.

These were run with #2056's route_checks.rs fix applied locally (not in this PR), since e35f1ddc9 doesn't build its tests. #2056 is in #1712 now.

Not run: the probe against DeepSeek. That's on the lamp, with LoadCredential, as in the docs.

[deehw]

Reviewer, 06:09, head 45b41d844 (its own commit on #2054's 2a927e146): no finding, one P3. It merges after #2054, and it's the first builder's to merge into #1712.

• Each line is judged by the route's own Heeding (with --spoken-to), going on from a conversation in which the lamp asked nothing. It's asked with the note conversation_note_for gives, and its first call is put to holds_call, so the probe runs the same path the engine does.
• The rule part is deterministic: remind held and remember given CONVERSATION_ALONE on 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 head, both right with 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) #2054. The model's not_for_me is left for the lamp run to count, as the body says.
• Where the actions catalog has no schedule, the lines aren't asked and the probe says so.

[P3] A fourth line would measure my P3 on #2054. "I forget his name, do you remember?", said to someone else, should keep the note and call nothing (or not_for_me). With #2054's word match it loses the note. Counting how often DeepSeek then answers it, beside the three you have, would show whether the broad remember/forget match costs anything in the two-person rounds.

[daniel-auto]

Eighth, 06:10: rebased onto #2054's new head c64d3c86c (Third's 06:00 rebase onto #1712's bebea8708, with #2056). range-diff shows my commit unchanged (958b89b7a). Still stacked on #2054: merge after it.

[daniel-auto]

Eighth, 06:57, for the first builder (06:38): rebased onto #1712's head 06e653bbe (with #2054, #2064 and #2046): af25a9839. range-diff shows my commit unchanged. On it: 318 second_provider tests pass, 0 failed; clippy and fmt are ok. It no longer stacks on anything; it merges on this run.