Context-threshold auto-reset scheduler — reset a daemon before it has to compact mid-task

MERGED to master on MAIN by orchestrator daemon 10 (proc 996). Codex 993 (wt2) built commit 7ec7540 on flower/context-threshold-reset; merged --no-ff as fce0ee9. Migration 2026_07_03_064500_add_reset_pending_to_daemon_agents_table applied on MAIN (reset_pending tinyint + reset_pending_at + reset_pending_context_tokens int, indexed) + config:clear. Tests green on MAIN (CoordinationQueueTest + DaemonRosterToolsTest: 18/18, 345 assertions; worktree full suite 629 pass / 1 skip, Pint clean). No horizon:terminate needed — new code runs in the checkin/MCP path, not Horizon jobs. Design delivered: (1) config flower.reset.context_threshold_tokens (400k default, env FLOWER_RESET_CONTEXT_THRESHOLD_TOKENS); (2) checkin flags/clears reset_pending on context crossing threshold (idempotent, audits on transition only, reads live linked-session usage); (3) hold-new-work gating in CoordinationQueue.signals()+claim() — pending targeted work held for reset_pending daemons while in-flight (claimed) + reset-lifecycle signals (reset/successor_ready) stay claimable; (4) AgentConventions line instructing daemons to finish beneficial work, stop pulling new, then daemon_request_reset. Reset stays decoupled: checkin only sets the flag; the daemon triggers the #76 make-before-break reset. Roster UI shows a reset-pending badge.

Operator ask (2026-07-03). The auto-trigger POLICY on top of #76 self-reset. Monitor each daemon's live context (context_used / context_percent, already on the roster via session_usage). When a daemon crosses a CONSERVATIVE threshold (config, e.g. 400k tokens — well before compaction), begin a graceful reset via #76 — but importantly: (1) the daemon first FINISHES the in-flight work where keeping context is beneficial (its current pipeline tasks); (2) it then HOLDS new work — stops pulling/accepting new tasks/signals — until the reset happens, so the reset is NOT perpetually deferred by incoming work; (3) then it triggers the #76 make-before-break self-reset (fresh session, clean context, current palette). The conservative threshold gives runway to finish the pipeline without compacting mid-task. Deliverables: config `reset.context_threshold_tokens` (+ maybe a percent variant); detection (on checkin, compare context_used vs threshold → set a `reset_pending` flag / enqueue the #76 reset after current work); the "hold new work until reset" gating on the daemon (finish current, stop pulling new, then reset — no perpetual deferral); AgentConventions for the daemon behavior. DEPENDS ON #76 (the reset mechanism). Scope the whole thing (operator: full setup, debug/pivot in dev if needed).