Feedback #53: recall_roster: add a compact/no-audit mode — per-daemon meta.audit arrays (100+ entries) bloat every polling daemon's context on each heartbeat check

Merged to master by flower-orchestrator (daemon 29). Merge commit 76c0536 (worker commit 3731712, branch flower/137-192-roster-slimming, off master @0289f3a — clean --no-ff, no conflicts). Adversarial review PASS: daemonPayload() untouched so top-level winddown/reset/compaction flags survive the lean default; only meta.audit history is trimmed (to last 3 via flower.daemons.roster_audit_tail) with meta.audit_count preserved; include_audit=true restores full trail. Full suite green on MAIN: 947 tests / 945 passed / 2 skipped / 0 failed. NOTE: recall_roster is an MCP tool — the leaner shape goes live in agent sessions only after the flower MCP server next reloads (existing sessions keep the old shape until then; expected).

Done on branch flower/137-192-roster-slimming (commit 3731712), designed together with #192 since both touch RecallRosterTool + DaemonRosterService. #137 (compact/no-audit mode): recall_roster now trims each daemon's meta.audit checkin trail to the last N entries by DEFAULT (config `flower.daemons.roster_audit_tail`, default 3), preserving the true length as `meta.audit_count`. New `include_audit=true` flag restores the full trail (and sets filters.audit_tail=null). This kills the 100+-entry-per-daemon bloat on the hot polling path. The lean shape is the default (the point of the brief); the escape hatch is opt-in. Response also gained a `filters` block echoing {include_retired, include_audit, hidden_retired, audit_tail} so a poller knows what was suppressed. Files: app/Mcp/Tools/RecallRosterTool.php (schema+handle), app/Services/DaemonRosterService.php (roster() + compactAuditPayload/rosterFilters/rosterAuditTail helpers), config/flower.php. Tests added in tests/Feature/Mcp/DaemonRosterToolsTest.php (audit trim default + include_audit full). Full suite green (946 passed, 1 pre-existing skip); pint clean on changed files. Not merged — left for orchestrator review.

Feedback #53 Authority: OPERATOR APPROVAL REQUIRED Funnel: A - operator approved brief Gate: this brief must not be dispatchable until the operator approves it. Kind: idea Source: flower-refine Summary: recall_roster: add a compact/no-audit mode — per-daemon meta.audit arrays (100+ entries) bloat every polling daemon's context on each heartbeat check Detail: Polling daemons (orchestrator/ops/refine) call recall_roster each cycle to check winddown_state / reset_state / needs_compaction, but the response embeds each daemon's full meta.audit checkin history (~100-150+ entries per daemon). For flower's roster (5 daemon rows) that's a very large payload every ~13 min, which measurably inflates the polling daemon's OWN context — ironically pushing it toward the compaction the check is meant to help avoid (observed: flower-refine context climbed several % per tick largely from roster pulls). Suggestion: an opt-in compact mode — e.g. `include_audit=false`, a `fields=` selector, or simply drop meta.audit by default and expose it via a separate detail call — returning just id/role/actor_ref/liveness/context_percent/winddown_state/reset_state/needs_compaction. That's all a polling daemon needs for the safety check. Would cut routine roster-poll cost dramatically across all daemons and projects. Ties into brief #97 (daemon charter/convention review). Context JSON: { "tool": "recall_roster", "impact": "inflates polling daemon context", "project": "flower", "observed": "meta.audit ~150 entries per daemon; large payload per heartbeat poll" }