Heartbeat poll and Persona turn durations measured on Yggdrasil, the busy
report, the dt cap, what the vote release does and does not promise, the
rerun table (27th by rank), and the crossing bound stated by the rows it
admits.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Conditional integration takes the caller's single saturation report (the
last beat was the floor, or launched nothing because every eligible Mind
was busy) instead of estimating the high limit from its own output; the
slow limit is judged on the integral's own answer, so the vote carry no
longer masks it. A call after a long silence integrates MAX_STEP_MS, not
the outage. The sim gains the swarm's throughput ceiling: per-Mind busy
time from the measured turn duration, a beat that finds every Mind busy
launches nothing, and the controller is told; scenarios cheap-then-x3 and
one-slow-mind exercise it; the band-crossing bound is per scenario.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Replaces the integrate-only-on-advance step: it changed the sampling the gains were tuned against for no gain over an age horizon, and left a replayed reading steering for as long as its age stayed inside it.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The integral no longer back-calculates to the voted rate (unbounded, it parked
the controller near the vote's burn after release): nothing integrates while
voted, and the gap between the voted rate (inside the beat bounds) and the
controller's own answer is a carry that fades over a poll period after
release. A reading whose observed_at has not advanced is not evidence: the
integral steps only on an advanced reading, at most one longest beat. Two
resets_at within a minute are one window.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Deletes first: the PostToolBatch and PreCompact hooks, --settings, the inject subcommand, the on-disk spool and its rename claim, close_past as a standing rule, close_told and the result-timed brief closure, Event::Told/PreCompact, Action::CloseBriefs, FLUSH_MARGIN_S.
Then: drive is the only writer into her session (start, tells, flush). decide counts lines written and lines echoed (--replay-user-messages) and derives idle from them; the flush is answered at the first result after its echo, folded or queued; an error before any answer or a blocked result closes at once; rotation and the idle cache end ask the same flush with a why; the grace is clamped so the signal ladder ends before the cap. DISABLE_COMPACT=1 replaces the PreCompact block. Briefs are follow-ups owned by self-run and sourced from a Persona turn; her Self resolves each she answers, the actuator closes only what she left open at an idle end. inbox::backlog lists, and never closes, the one-time sweep.
Tests: the committed 2.1.312 streams replay through read_usage and decide; the unit tables, a ladder sweep and an event-order enumeration pin the rules.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A mutation run over the memory diff left these unpinned: the asks the schema
offers, next_sequence counting per subject, learned_in cut at the Short bound
on a character boundary, other_senders excluding her own and repeats, and
memory_people listing the operator once. The place cut loses a redundant
boundary test: pop removes whole characters.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Lands the parked closure (eureka-body/self-session-live-closure-parked 13de7f0) on the Recorded outcome of ruling brief-closure-recorded.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>