fix(orch): exempt managers from per-target worker capacity

[nathanwhit]

Symptom

On objective d3bdf790 (PR #35461), a user PR comment went unanswered for >5 min. Events showed pr_feedback_ingested but no followup_spawned after it — unlike every prior feedback round, which spawns a follow-up within ~2 min. The comment only got a follow-up ~9 min later, and only because a local slot happened to free.

Root cause

SyncPRFeedback runs IngestFeedback → ProcessFeedback → spawnPRFollowup, whose first step is SelectTarget(TargetRequest{}). It returned ErrNoTarget because every target was unplaceable:

• local (cap 4): available_sessions = 0 — full of three idle interactive managers, each holding a per-target slot for its whole life.
• Vultr (cap 16, 15 free): load_per_core = 1.89 ≥ the default -max-load-per-core 1.5, so the load gate skipped it.

spawnPRFollowup returns that error without spawning, marking the feedback handled, or emitting any event — so it is silently swallowed and retried invisibly every 1-min monitor tick (ingest dedups → no event; spawn keeps failing → no event). Nothing in the events DB or orcha stdout surfaces it.

The design gap

orcha already exempts managers from the global worker-concurrency cap (Store.CountActiveWorkerSessions skips RoleManager; Scheduler.Tick lets managers bypass the budget), precisely because "managers sit idle most of their life ... counting them starves new work." But it left them charged against per-target capacity (available_sessions). So sum(capacity) (here 4+16=20) is a hard ceiling on resident managers — 20 managers ⇒ zero worker slots ⇒ fleet deadlock.

Fix

Extend the existing global-cap exemption to per-target capacity:

• TargetRequest.IgnoreWorkerCapacity (set by targetRequestFor for managers) makes SelectTarget skip the AvailableSessions <= 0 check.
• PlaceSession / releaseTargetSlot skip the claim/release for managers — both keyed on a single sessionExemptFromCapacity(sess) predicate (role == manager) so claim and release can never disagree.
• Managers stay bounded by the load gate and the per-objective respawn limit.

available_sessions is an incrementally-maintained counter, so make it self-healing: Store.ReconcileTargetSlots rebuilds it from live non-terminal, non-manager occupancy, called once in RecoverInterrupted at startup (before the scheduler's first tick). This both heals crash drift and frees the slots currently-running managers claimed under the old accounting when this deploys.

Tests

New internal/orch/capacity_test.go:

• manager places on a full target while a worker is refused
• placing/releasing a manager doesn't move available_sessions; a worker does
• ReconcileTargetSlots excludes managers + terminal rows, heals a drifted counter, and is idempotent

go test ./... and go vet ./internal/... pass.

Not in this PR (follow-up)

The silent ErrNoTarget swallow in spawnPRFollowup — a placement-starved follow-up should emit a throttled audit event so it's visible instead of invisible. Flagged for a small follow-up.