CONTEXT

Child of 8894dd80 (model adjustments for the web UI). See that parent ticket for motivation, non-gaps, out-of-scope list, and parallelization map. This ticket is narrowly scoped to Gap 2.

================================================================ THE CHANGE

handle_get_turn in src/mcp.rs currently returns:

{ "message": "Turn N at simulation_time T.", "turn_ref": "turn_NNNNNN", "state": }

Add an optional include_events: bool arg. When true, the response gains a new top-level events field containing every audit event whose turn field equals this turn number, in _seq order (ascending — chronological).

When include_events is absent or false, the response is byte-identical to today's. This change is strictly additive for existing callers.

================================================================ WHAT'S IN THE EVENTS ARRAY

All events for turn == N, regardless of attempt_status. This means failed-attempt events are included too — adjudication_rejected events, and the attempt_failed terminator when an attempt did not commit. The view for the turn page wants the full story, including retries.

Do NOT filter by attempt_status. Do NOT apply the include_failed default-false that get_events uses — get_turn(include_events=true) is asking for "everything that happened while the system was trying to produce this turn," which by definition includes the failed attempts before (or leading up to) the committed one.

Order by _seq ascending. Do not re-sort by event_type or attempt_status or anything else — _seq is the canonical ordering.

================================================================ IMPLEMENTATION NOTES

The audit log is at events.jsonl under the world dir; accessible via AuditLog in src/persistence.rs. Whatever read path get_events currently uses is the same one this handler wants. Single pass, filter by turn == N, collect.

Turn number comes from parsing the turn_ref (strip "turn_" prefix, parse as u64) — same trick the existing handler uses.

Edge case: if turn_ref refers to a turn that exists in the turn chain but has no audit events (shouldn't happen for any turn >= 1, but turn 0 is the seed and emits no events), return events: [] rather than erroring.

================================================================ WIRE CONTRACT

Request: { "world_slug": "ant-verify", "turn_ref": "turn_000001", // OR "turn": 1, // existing turn/turn_ref "include_events": true // NEW, optional, default false }

Response when include_events: true: { "message": "Turn 1 at simulation_time ...", "turn_ref": "turn_000001", "state": { ... }, "events": [ { "_seq": 1, "turn": 1, "event_type": "perception_emitted", ... }, { "_seq": 2, "turn": 1, "event_type": "intent_formed", ... }, { "_seq": 3, "turn": 1, "event_type": "intent_adjudicated", ... }, { "_seq": 4, "turn": 1, "event_type": "turn_complete", ... } ] }

Response when include_events is absent or false: unchanged from today.

================================================================ TESTING

Add test: seed a world, run a turn, call get_turn(turn=1) without the flag — assert no events field in response. Call again with include_events: true — assert events is a non-empty array, events are in ascending _seq, every event's turn field equals 1.

If the suite has a test that induces an adjudication retry (even via a mocked router; but per prior guidance tests against the live router are fine), assert that adjudication_rejected events appear in the array for the relevant turn.

================================================================ ACCEPTANCE

• cargo build clean
• cargo test green
• Response without include_events is byte-identical to current
• Response with include_events: true contains an events array ordered by _seq, including failed-attempt events
• Tool manifest description for get_turn updated to mention include_events
• No kernel change
• No persistence-layer change (just a new read path)
• No other handler modified

================================================================ SCOPE DISCIPLINE

• Do NOT change the default behavior. include_events must default to false.
• Do NOT filter failed attempts out when include_events is true.
• Do NOT add a separate include_failed knob. One flag, one behavior.
• Do NOT modify get_events. That handler keeps its existing default.

Implemented, committed, merged, deployed, and smoke-verified in production.

What changed (src/mcp.rs only):

• handle_get_turn: strictly additive include_events: bool arg (default false via .as_bool().unwrap_or(false) — matches the file's 15+ other bool-arg call sites). When absent or false, response is byte-identical to pre-flag shape. When true, response gains a top-level events array containing every audit event whose turn field equals this turn, ordered by _seq ascending.
• All events included regardless of attempt_status — committed path events plus any adjudication_rejected / attempt_failed from retries. No include_failed-style knob; one flag, one behavior, per spec.
• Read path: single rt.audit.read() pass + inline filter by turn. _seq tagging matches the 1-based convention EventQuery::run uses, so round-trips are consistent across handlers.
• Tool manifest description updated with the new flag, its default, and the "includes failed-attempt events" semantics.

Unchanged: get_events default behavior; kernel; persistence write path; any other handler.

Receipts:

• Commit: 0fbef7f feat(mcp): get_turn accepts include_events flag
• Merge: 9e7fc45 on main (ort strategy, clean)
• cargo test --lib: 244 passed in worktree; 279 passed in merged main.
• cargo test --test phase0: 14 passed.
• cargo test --test ant_scenario: 3 passed in merged main (regression — this ticket didn't add live-router coverage since the unit suite exercises the handler directly).
• New unit tests in mcp::tests: get_turn_without_flag_omits_events_key (both absent and explicit false cases), get_turn_include_events_returns_only_matching_turn_events_asc_by_seq (multi-turn fixture including distractor on turn 5 and adjudication_rejected+attempt_failed on turn 0 — asserts seq sequence [1,2,3,5], exactly 4 events, every event's turn==0), get_turn_include_events_turn_zero_with_empty_audit_returns_empty_array, get_turn_unknown_turn_still_errors_with_include_events — all pass.

Deploy + production smoke:

• bash k8s/deploy.sh clean; pod chukwa-69bf5755d7-tddh5 Running 1/1.
• GET /v1/tooling/manifest confirms get_turn.inputSchema.properties now includes include_events.
• Created world smoke-rollup, ran turn 1 (committed 29.6s).
• get_turn(world_slug=smoke-rollup, turn=1, include_events=true) response included an events array with 4 events in _seq order [1,2,3,4]: perception_emitted → intent_formed → intent_adjudicated → turn_complete. Every event's turn field equals 1, every event carries world_slug: "smoke-rollup".
• World deleted after verification.

Child of 8894dd80. I am not confirming — over to you.