fix(agents): senior-pm delegation reply shape — routing floor 0.75→0.9 (G#393)

[psdjungpulzze]

Goal

G#393 (6eeb182f-e1ae-45f7-89a3-f0ccd883aa47): Strengthen senior-pm routing adherence for roadmap/discovery/drift/phase-sequencing intents; raise the eval capability floor from 0.75 toward 0.9. Follow-up to G#352 (PR #1751) and G#350 (PR #1736).

Root cause (measured, not guessed)

Targeted live triage (orch-002/003/005/019, 5 samples each, judge reasoning read) showed the routing family failing on stall replies, not wrong-specialist choice:

• orch-002: 0/5 — every sample was "I'll pull the current project state first, then route this to the right specialist" (no specialist named, no delegation visible). Induced by the prompt's session-start project_query habit (rule 1).
• orch-003: 0/5 — same unnamed-specialist stall; one sample fabricated <function_calls> syntax and invented the specialist's entire deliverable.
• orch-005 / orch-019: 4/5 each — flips were the same stall pattern.

The eval judge sees only the agent output (never the scenario input), so a delegation that doesn't name its specialist is unscoreable — and in production the same stall is simply a worse reply.

Fix (src/server/agents/configs/senior-pm.ts, additive — G#352 invariants untouched)

• New "Delegation reply shape" block in ROUTING_RULES: the first reply IS the delegation — name the exact specialist, one-line reason, *[Delegated to …]* marker; bans stall replies, pre-routing project_query, and fabricated delegation narration / invented specialist output.
• Tie-breakers for overlapping intents: sequencing-over-time vs value-ranking (roadmap vs prioritizer), metric-drop → discovery, baseline-vs-current scope → drift analyst even when a phase is named.
• Rule 1 gains an explicit routing exception (delegate first; the specialist gathers its own context).

Validation (paid live evals, CI parity)

• Targeted post-fix rerun: 20/20 samples across the four scenarios (was 9/20).
• 3 full live runs at EVAL_SAMPLES=5: capability 1.00 / 1.00 / 0.95, canaries 100% in all three. orch-002/003/005/019: zero flips across all runs (done-when tolerated one).
• The single flip (orch-008, majority 2/5, run 3 only) is a judge-blindness artifact: the scenario's systemContext supplies "Phase 2 is active", but the judge can't see inputs, so restating it reads as data invention. Documented in the floor note; recommended follow-up: pass scenario input to the judge prompt.

Floor

evals/capability-floors.json senior-pm-orchestrator 0.75 → 0.9 (the default — fully reached), note rewritten with the evidence trail. Measured low 0.95 leaves one-flip headroom on the 20-scenario suite.

Agent-maintenance note

This change is the agent config update (prompt-only; no tools added/removed, no TOOL_NAMES or config-sync changes needed).

Tests

• config-sync / capability-floors / eval-agents-resolver suites: 49 + 7 pass.
• npm run typecheck clean.