[Scheduler] Add fair admission control, bounded backpressure and equivalent-work inference deduplication

[wesleysimplicio]

Parent: #674

Objective

Prevent worker/process explosion and duplicated local inference when many issues or subissues arrive at once. Admit work according to observed capacity, preserve fairness and share only safe equivalent in-flight inference.

Implementation steps

1. Model separate limits for runnable tasks, active workers, inference requests, backend slots and memory budget.
2. Add bounded queues with explicit saturation behavior.
3. Implement weighted fair queueing across repositories/sessions/priorities with aging.
4. Derive an equivalence key from model/backend identity, stable-prefix generation, canonical request hash, tool-registry generation and safety scope.
5. Coalesce only side-effect-free inference requests with identical authority/context boundaries.
6. Never deduplicate tool execution, mutations, approvals or user-private contexts.
7. Fan out the same immutable completion/receipt to equivalent waiters with separate correlation ids.
8. Propagate cancellation correctly: one waiter cancel must not cancel shared work needed by others.
9. Add retry budgets with jitter and circuit breaking; prevent retry storms.
10. Emit queue depth, wait time, admission decision, dedup hit/miss, fairness and saturation metrics.

Tests

• Queue capacity and deterministic rejection/defer behavior.
• Fairness/starvation simulation with hot and cold sessions.
• Dedup positive/negative matrix including one-byte prompt, authority and tool-generation differences.
• Cancellation with one/all shared waiters.
• Backend outage/circuit breaker and bounded recovery.
• Load/property tests proving queue and worker counts never exceed caps.
• Security test proving different users/scopes cannot share inference.
• E2E burst workload comparing duplicate calls and wall time.

Acceptance criteria

• No unbounded queue or automatic worker multiplication exists.
• Equivalent inference is coalesced only before side effects.
• Security/authority scope is part of the equivalence key.
• Every waiter receives its own trace correlation and truthful shared-execution reference.
• Fairness tests demonstrate bounded starvation.
• Saturation degrades through explicit defer/block/fallback states.
• Retry storms are mechanically prevented.

[wesleysimplicio]

🔁 Simplicio Loop — status da execução

Campo	Valor
Estado	CLAIMED
Issue	#676
Run / attempt	wave1-agent676 / wave1-agent676
Agente	Codex — issue-676
Lease / fence	wave1-agent676 / fence-2-9ef2f18174924bb9b487920fa84b43d9
Branch / worktree	codex/issue-676-fair-admission / C:\Users\Z0059V7A\m\repos\work\simplicio-loop-issue-676
Base	origin/main @ fa237214a9d420fa0cf54c4c548add7d7164c88e9

Objetivo e escopo

Implementar somente o issue #676: fair admission, bounded backpressure e safe equivalent-work inference deduplication. Não alterar código dos issues #675, #677, #678, #681, #685 ou #686.

Status

Lease de backlog reivindicada e retida; survey/planejamento e validação serão publicados como evidência. Blockers obrigatórios de autorização, source-drift ou verificação não serão atravessados.

[wesleysimplicio]

Progress receipt: PR #698 merged as