Track structured Auto Memory rollout readiness on main #12947

Goal

Track the remaining correctness, effectiveness, and validation work for structured Auto Memory on main before considering broader rollout.

This is the memory-specific closeout tracker for #10151 and a child track of the broader token-governance umbrella #12028. Structured recall remains intentionally opt-in and disabled by default while this work is incomplete. Completing this checklist enables a separate default-on decision; it does not require or pre-commit that decision.

Baseline reviewed for this tracker: main at 92f4d4f6455bf02a41e0d856afe63b79108f936f.

Landed foundation

• feat(memory): extract structured scan and bounded retrieval #12726 — structured scan and bounded retrieval primitives.
• feat(memory): extract metadata migration and writer compatibility #12757 — metadata migration and writer compatibility.
• feat(memory): add structured on-demand recall #10183 — structured tree/focused recall, on-demand search_memory, activation, and User Dream integration.

Remaining correctness work

• fix(memory): schedule legacy metadata migration after tool-completing turns #12929 — schedule legacy metadata migration after a tool-completing interaction. The current fix is included in fix(memory): preserve migration progress when index rebuild fails #12913.
• Managed auto-memory extraction is skipped after tool-completing CLI turns #12938 — schedule extraction and Dream exactly once after a tool-completing CLI interaction. The current fix is included in fix(memory): preserve migration progress when index rebuild fails #12913 at 7a156ee56a, using the originating interaction type rather than the final ToolResult type.
• fix(memory): preserve migration progress when index rebuild fails #12913 — preserve committed migration progress and recovery state when index rebuild fails, in addition to the fix(memory): schedule legacy metadata migration after tool-completing turns #12929 scheduling fix.

Remaining hardening work

• fix(memory): tighten structured recall contracts #12916 — close the structured-recall contract, bounded-work, validation, and regression-witness gaps deferred from feat(memory): add structured on-demand recall #10183.
• fix(memory): preserve task lifecycle signals #12917 — close User Dream telemetry and cancellation lifecycle gaps deferred from feat(memory): add structured on-demand recall #10183.
• Revalidate follow-up(memory): resolve non-blocking review debt after #10183 #12853 after fix(memory): tighten structured recall contracts #12916 and fix(memory): preserve task lifecycle signals #12917 merge. Keep genuinely non-blocking P3 items deferred instead of treating the entire debt list as a rollout blocker.

Effectiveness validation

• Close Validate the recall initial-turn budget against production delivery telemetry #8998 with a controlled selector ablation. Measure unique selector deliveries, overlap with deterministic fast recall, already_delivered, no_safe_delivery_point, answer quality, and the extra selector/model requests. Do not use Validate the recall initial-turn budget against production delivery telemetry #8998 as the umbrella tracker.
• Rerun the matched end-to-end matrix after the correctness fixes land: legacy-only corpus, fully structured corpus, metadata-only answer, body fetch, tool-using interaction, unrelated query, and compression/refetch.
• Report main-model uncached input, cache-read input, output, request count, selector usage, extraction usage, Dream usage, and migration usage separately. Do not present a lower whole-process total as an optimization when a background task was skipped.
• Recheck tail latency. The earlier 36-case result improved the median but regressed P95 (191,893 ms to 276,311 ms); the final main implementation needs a matched rerun before rollout.
• Confirm answer and recall quality do not regress while token reduction remains meaningful on the final implementation. The corrected three-case matched smoke reduced main-model gross input from 73,732 to 62,571 tokens (-15.14%) with 3/3 answers passing. The observed whole-process -27.20% is not a stable claim because extractor request counts differed between arms; this remains directional evidence, not the final acceptance result.

Rollout gate

Keep memory.enableStructuredRecall opt-in until:

• the correctness items above are fixed;
• legacy and structured memory paths both pass the end-to-end matrix;
• the measured token result separates foreground savings from background cost;
• selector contribution and tail latency have a clear disposition; and
• no tool-using interaction silently skips required memory lifecycle work.

Default-on remains a product and rollout decision after these gates pass, not an automatic consequence of closing this tracker.

Relationship to other issues

• tracking(core): non-conversation context token governance #12028 is the overall non-conversation context and token-governance umbrella.
• Improve Auto Memory with structured recall and lossless migration #10151 is the original design and feature umbrella.
• This issue is the current-main closeout and rollout-readiness checklist.
• Validate the recall initial-turn budget against production delivery telemetry #8998 tracks only selector timing and marginal value.
• follow-up(memory): resolve non-blocking review debt after #10183 #12853 tracks deferred non-blocking review debt.
• fix(memory): schedule legacy metadata migration after tool-completing turns #12929 and Managed auto-memory extraction is skipped after tool-completing CLI turns #12938 track concrete interaction-lifecycle bugs.