Spec Kitty

Research: composition-advance-alignment

R-1 Grounding

• Decision: close the gap by routing the composition advance through engine.plan_advance / commit_advance (operator ruling: a specified behaviour change).
• Rationale: GROUNDING.md §"#2562 design-spike finding" shows that the adapter's planner runs only apply_result + plan_next. A composition-backed advance therefore records no SignificanceEvaluated event and no raci: record.
• Alternatives considered:
• Keep the parallel planner and add the two calls to it. Rejected: that keeps two planning authorities, which goes against the charter's single-canonical-authority principle.
• Behaviour-preserving dedup only. Rejected by the operator ruling.

R-2 Post-spec adversarial squad (2026-10-06, advisory)

Two lenses ran: consumer impact (reviewer-renata profile) and design invariants (architect-alphonso profile). Each finding's disposition:

#	Lens	Severity	Finding	Disposition
1	both	BLOCKER/MAJOR	The adapter unit tests (test_bridge_engine.py:341-640, test_bridge_decision_log_flush.py:359-400) stub _planner.plan_next with an object() template. plan_advance reads neither, so these tests crash.	Folded. Spec US3 now allows re-seaming them onto a real minimal frozen template with every assertion kept. Each re-seam is listed in the PR.
2	design	MAJOR	A complete_run(emit) hook could drop or duplicate MissionRunCompleted.	Folded. The hook becomes `before_run_completed: Callable[[], None] \
3	design	MAJOR	A malformed significance block now blocks a composition advance (EDGE-003).	Folded. It is an edge case in the spec and is named in the FR-012 changelog entry. This matches the engine path.
4	design	MINOR	plan_composition_advance would forward to plan_advance (C-003).	Folded. It is deleted together with CompositionAdvancePlan, and the bridge calls _engine_adapter.plan_advance(..., "success").
5	design	MINOR	Symbols become dead: the apply_result wrapper, _mark_step_completed, _live_template_path and the payload imports.	Folded. They are deleted in IC-03. The _append_event / _write_snapshot wrappers are kept (pinned by the six-wrapper test).
6	design	MINOR	The legacy path catches StaleAdvancePlan and falls back to runtime_next_step; that must not be copied here.	Folded. The FR-010 test asserts that runtime_next_step is never called.
7	consumer	MINOR	A run paused at a MEDIUM gate from before this change can mismatch: it holds approve/reject options and no significance record.	Recorded residual (spec edge cases). The composition path never re-polls a pending decision, so this mission does not reach that state.
8	consumer	MINOR	DecisionGitLog committed content changes for MEDIUM/LOW gates.	Folded. It is an edge case in the spec and is named in the changelog.
9	consumer	NOTE	Answered records gain raci_source, which nothing outside the engine reads. No contract changes (events, Decision, orchestrator-api CONTRACT_VERSION 1.10.0).	No action. C-005 holds, so no escalation to the orchestrator.
10	design	NOTE	Q1: StaleAdvancePlan cannot fire spuriously, because nothing writes state.json between plan and commit.	No action; the analysis is recorded.
11	design	NOTE	The legacy and composition strict gates differ on refusal: the composition path leaves NextStepAutoCompleted in the log, while the legacy path rolls back.	Out of scope (preserved deliberately); named under the PR's Deferred section.
12	design	NOTE	Emitter seeding could reuse plan.source.	Kept as _read_snapshot to preserve today's behaviour exactly; this is a NOTE only.
13	design	Q7	Proposed split of provide_decision_answer.	Adopted as the IC-01 helper map.

No finding needs an operator decision beyond the ruling. Both lenses confirm that C-005 holds.

R-3: Post-tasks squad (planner-priti lens, 2026-10-06, advisory)

There were 4 MAJOR findings, all folded:

Finding	Disposition
(1) Silently dead planner.plan_next patches	Repatch onto engine.plan_next, with a fake-ran assertion
(2) Retrospective order not pinned on the composition path	Order and strict-abort tests added to WP02
(3) The FR-010 red case was missing (C-004)	Committed red in the gate file
(4) The untracked gate file would block implement	The gate was committed before WP01

Minor findings 5-11 and 13-14 are folded into the WP02 prompt ("Post-tasks squad folds") and the WP01 DoD. Note 12 (non-blocking capture now runs after the snapshot write) is recorded for the PR. Note 15 needs no action: the gates are confirmed safe. Note 16 also needs no action: the historical ADR is left as is.

R-4: Pre-PR squad (2026-10-06, advisory)

Two lenses ran: correctness (debugger-debbie) and boundary/overlap (architect-alphonso). Neither found a BLOCKER or a MAJOR.

#	Lens	Severity	Finding	Disposition
1	correctness	MINOR	A failure while building the best-effort retrospective capture, after the commit, turned a terminal advance into a false blocked Decision.	Fixed at the root (runtime_bridge_retrospective). The callback is now built inside the policy's try. Red-first test in tests/runtime/test_bridge_retrospective.py.
2	correctness	MINOR	A strict-gate refusal leaves NextStepAutoCompleted (and now SignificanceEvaluated) in the run journal. The legacy path rolls back; the composition path does not.	Fixed at landing (operator ruling 2026-10-07). The guard runs before anything is appended or emitted, so a refusal writes nothing and a retry does not stack events.
3	correctness	NOTE	The strict-refusal operator message differs between the paths, and a stale plan re-dispatches the composed action on the next call.	Deferred / by design (FR-010). Named in the PR.
4	boundary	MINOR	The adapter's _append_event / _write_snapshot wrappers are dead, kept alive only by a pin test.	Fixed. Wrappers deleted. The pin covers reads and planner names, and a new check asserts the adapter wraps no writer.
5	boundary	MINOR	The shape gate missed SignificanceEvaluatedPayload and direct writer calls.	Fixed, with planted rows.
6	boundary	MINOR	The terminal predicate is restated in the adapter.	Deferred. It is low risk. An after_run_completed engine hook is named in the PR.
7	boundary	MINOR	The adapter docstring cited stale call-site lines.	Fixed.
8	boundary	MINOR	The no-bridge positive-presence check has no planted row.	Accepted: it is a positive check.
9	boundary	NOTE	Upgrade note for readers of decisions.events.jsonl.	Added to the CHANGELOG.
10	boundary	NOTE	engine.apply_result / existing_template_path could become private again.	Out of scope (#5834 surface).

R-5: Landing residuals (2026-10-07)

Ref	Residual	Disposition
#5853	spec-kitty next in query mode previews a MEDIUM audit gate as approve/reject, but the persisted request and the answer validator use decide_solo / open_stand_up / defer: query mode plans through plan_next on the live template and never evaluates significance. The engine path had this gap before this change and the composition path now shares it. Only operator-authored templates that declare significance: reach it.	Open follow-up; plan_advance applies a result, so it cannot be dropped into query mode as is.
#5854	The stale-plan check in commit_advance detects a run that moved between plan and commit but does not serialise concurrent writers. On a stale plan the engine path (_dn_advance_engine) re-plans through next_step and can complete a step this caller never ran; the composition path refuses with a blocked Decision.	Open follow-up (a run-dir lock, and a refusing engine path). The limits are stated in the commit_advance docstring and the changelog.
#5855	Re-polling a pending audit gate evaluates its significance again. A run paused at a MEDIUM gate before this change keeps its approve/reject request, so a re-poll records the MEDIUM band and a later approve is refused. Engine re-poll behaviour that predates this change.	Open follow-up. Stated in the changelog Upgrade Notes: answer such a gate before polling again.