0105: Run semantic passes only for actionable work

Status: accepted (2026-08-23) · Scope: semantic service

Context

ADR 0102 starts semantic reading with the dashboard. In the first live run, the service completed 1,378 sweeps and 613 passes while making 60 model calls. Many passes were useful incremental projections, but the daemon also started a pass on every sweep whenever its manifest had pending work. A pending generation can remain ineligible for 30 seconds after its latest model call, so those passes could not call a model or settle the pending item.

The scheduler already owns the debounce, per-generation cooldown, and rolling hourly cap. The daemon should use those rules before starting a pass instead of discovering the same ineligibility deeper in the reader.

Alternatives

Keep running every pending sweep. Rejected because it creates repeated manifest reads, projection work, and status writes without advancing the read model.

Increase the global wake or debounce interval. Rejected because it would delay newly actionable ledger evidence and treat all work as if it shared one cooldown.

Lower the hourly model-call cap. Rejected for this change because the cap limits spend, not orchestration churn. Token accounting should provide evidence before Harnery changes the budget.

Decision

The semantic scheduler exposes a pending-pass predicate. A pending-triggered pass is due only when an item has completed its debounce and either:

• at least one matured generation is outside its per-generation call cooldown;
• or the rolling hourly cap has no capacity, in which case one pass records deferred receipts and removes the matured items from the pending queue.

The daemon still runs for matured raw-ledger changes and eligible deferred work. The 60-call rolling ceiling, model routes, privacy filter, expression contract, evidence bounds, and deterministic Codec precedence do not change.

Result

Focused scheduler tests cover held generations, an eligible peer beside a held generation, pending debounce, and the deferred-receipt pass at the hourly cap. Daemon coverage confirms held-only pending work can sweep without starting a semantic pass. The existing semantic suites continue to pass with usage accounting enabled.