Assign owners and record decisions for pending architecture proof harnesses #210

Outcome

Every pending architecture proof obligation has a current owner and a recorded disposition: fulfilled by an identified binding harness, scheduled under an existing or bounded issue, or explicitly deferred with its limitation. A declared metric or invariant is never described as an enforced runtime gate merely because the architecture check passes.

Parent and boundary

Parent: #108 (G00). Owner: architecture/CI maintainers, with the owning subsystem reviewer.
This issue owns the proof-obligation inventory and decisions, not a blanket architecture refactor or a second implementation of L04, L14, or R00. Map existing work first: #33/#35/#37/#39/#40, #95/#96/#136, #94, #51, and #185 as applicable. Reuse active PR #208 for #33 and PR #202 for Compare; do not open parallel outcome branches.

Grounding

Reviewed integration SHA: 0704eef.

• architecture/invariants.yaml records the clock/RNG classification TODO and five declared D1-D5 obligations: canonical-JSON call-site guard, nonpublishing-preview guard, destination-envelope guard, publication/durable-completion state model, and named cross-surface CI fixtures.
• architecture/quality-budgets.yaml has seven declared/pending metric harnesses: search-latency, bleed-fixup-latency, resident-memory, parser-differential-agreement, packaging-size, sealed-rendering-channel-delta, and canonical-json-digest-stability.
• check-architecture.py requires a pending TODO for declared metrics; it does not measure them. D1-D5 already have behavioral/golden evidence; the declared guard TODOs are additional proof coverage, not evidence that the primitives are absent.
• Existing R00-08 — Gate performance and scale on benchmark identities #136 and L14-06 — Gate present-lane frame budgets on a pinned reference machine #185 own performance outcomes, and L04 owns governed correctness, but the issue suite has no complete row-level disposition for these pending obligations. Active Compare work adds another declared metric; reconcile it only after it lands.

Acceptance

• A reviewed table covers the clock/RNG TODO, D1-D5 TODOs, and every declared/pending quality metric at the execution SHA, including subsequently merged rows. Each row names its current harness/evidence, owner, live issue/PR or explicit deferral, and closure criterion.
• Reuse existing owners and tests. File a bounded child only where a distinct accepted harness task has no owner; do not recreate L04-01 — Reconcile registry and canonical plan contract #33, L04-08 — Implement forward rollback and cross-surface parity #40, R00-08 — Gate performance and scale on benchmark identities #136, L14-06 — Gate present-lane frame budgets on a pinned reference machine #185, or the independent-oracle/Medic gates.
• For an admitted enforcement change, record measured baseline and full fixture/tool/platform identity, bind the producer and comparator to CI/check-change, and show a deliberate violating measurement or fixture fails. Missing measurements remain unavailable or incomplete.
• Preserve explicit declared/deferred status when a harness is not admitted. Do not guess baselines, loosen fixed budgets, change sealed expected outputs, or promote a TODO to enforced without proof.
• A new schema or exported API needed by a proposed guard requires an architect checkpoint with concrete fixtures and migration behavior before implementation.
• Failure case: a passing structural architecture check with pending harnesses cannot certify those metrics, remove their limitations, or satisfy release admission.

Required proof

Row-level issue/harness read-back and reviewer decision. Any implementation runs its subsystem's binding lanes, architecture/catalog checks, and the normal check-change gate with one scoped fragment and evidence manifest. Runtime measurement requires a suitable existing build/runner; unavailable lanes must be recorded.

Handoff

Record the table, linked existing and new work, exact source SHA, implemented enforcement, and retained limitations. Include a 1-3 sentence quality summary.

Sources

Approved master roadmap §5 and §6.3-6.5; the pinned architecture/checker sources above. Follow-up gap review, 2026-10-04.

G00 implementation record, 2026-10-10

PR #262 publishes the row-level proof table. It covers clock/RNG, D1-D5, and all eight current pending quality metrics, including Compare. #202 and #208 are now merged provenance rather than active outcome branches. D5 has binding fixtures and a comparator; the architecture declaration remains declared. Other supplemental guards and measurements retain explicit owners, deferrals, limitations and closure criteria. No baseline or sealed output changed, and no new metric enforcement was admitted.

Implementation head: 78f9c0bce909bca7f95cc5b6c701043e61332087. The required check-change gate passed using the existing build with identical plugin ABI inputs. 117 Python cases and the mapped ABI test passed; one optional PdfTool runtime case skipped. Hosted CodeQL checks passed at this head; the native CI build is still running. Full package qualification remains pending. This record does not grant release admission.