v0.7.0 Review Synthesis — Fix Dispatch Backlog

Synthesized from 6 parallel reviewers (security/correctness/perf/API-UX/docs/coverage) Base commit: 0536e96 (feat/v0.7.0-grand-slam HEAD with CHANGELOG + ship-readiness fixes) Reviewer baseline: ef92bd7 / 64528b1 (pre-ship-readiness reconciled trunk) Issue: #767 Date: 2026-05-15

Reviewer source counts (raw findings)

Prompt header said “103 findings”; actual count from re-reading the six docs is 111 (likely 103 was based on a pre-finalization head; net total higher by 8). All 111 are absorbed below or classified as INFO / accepted-debt.

Severity summary (after dedup + cluster)

• CRITICAL: 4 fixes (dedupe’d from 6 raw CRIT findings — one already resolved, one merged)
• HIGH: 13 fixes (dedupe’d from 28 raw HIGH findings)
• MEDIUM: 14 fixes (dedupe’d from 43 raw MED)
• LOW: 10 fixes (dedupe’d from 27 raw LOW)
• INFO / Already-fixed / accepted-debt: 14 (dedupe’d from 7 INFO + cross-lens dupes)
• Cluster count: 12 (A–L)

Known dedupes applied

Fix dispatch clusters

Each cluster is a self-contained fix-agent prompt. Clusters are designed to be parallel-safe unless the Parallel-safe with / Dependencies rows say otherwise.

CLUSTER A — Form 4 fact-provenance correctness + atomisation idempotency

• Scope:
  1. Fix UTF-8 panic in compute_atom_span: replace start.saturating_add(1) with a codepoint-aware advance (use body.char_indices() or floor_char_boundary).
  2. Distinguish QueryReturnedNoRows from real errors in read_atomised_into — propagate real errors via ? so the idempotency guard doesn’t double-atomise on lock-timeout.
  3. Replace unwrap_or_default() / .ok().and_then(serde_json::from_str) silent drops in row_to_memory for citations / source_span / confidence_signals with the existing metadata-style tracing::warn!("corrupt … defaulting to …", row_id) pattern + a counter metric.
  4. Push source_uri_prefix filter into recall SQL WHERE clause so idx_memories_source_uri is actually consulted (currently dead code).
  5. Tighten validate_source_span to require end <= body.len() and char-boundary checks.
  6. Tests: add compute_atom_span_paraphrase_fallback_returns_none, compute_atom_span_multibyte_utf8_stays_on_char_boundary, recall_with_source_uri_prefix_uses_idx_memories_source_uri (EXPLAIN QUERY PLAN check), and a row_to_memory_corrupt_citations_logs_and_returns_default.
• Files (absolute):
  • /Users/fate/v07/synth-fix/src/atomisation/mod.rs (compute_atom_span, read_atomised_into)
  • /Users/fate/v07/synth-fix/src/storage/mod.rs (row_to_memory)
  • /Users/fate/v07/synth-fix/src/validate.rs (validate_source_span)
  • /Users/fate/v07/synth-fix/src/cli/recall.rs and src/storage/mod.rs recall SQL site (push source_uri_prefix into WHERE)
  • /Users/fate/v07/synth-fix/tests/form_4_provenance.rs, /Users/fate/v07/synth-fix/tests/atomisation.rs
• Findings absorbed: COR-1 CRITICAL, COR-2 HIGH, COR-3 HIGH, COR-7 MED, PERF-3 HIGH, COV-3 HIGH
• Acceptance tests required: four new tests listed above; CI gate green.
• LOE: ~1 session (3–4 hours)
• Parallel-safe with: B, C, D, E, G, H, I, J
• Dependencies: none

CLUSTER B — Form 1 synthesis security + verdict-application correctness + prompt-injection guard

• Scope:
  1. SEC-1 fix: re-run the K9 permission gate on every verb:"delete" candidate before applying; refuse verdict batches that delete > N (default N=1) without an explicit K10 approval flow; wrap incoming title/content into a <USER_CONTENT>...</USER_CONTENT> envelope in the synthesis prompt so the system prompt tells the model to treat as opaque.
  2. COR-5 fix: decide policy on multi-update verdicts (recommend honour all updates in sequence; if not, emit tracing::warn! per dropped update and adjust synthesis_decisions.update_dropped).
  3. COR-6 fix: surface synthesis failure in the response envelope (synthesis_failed: true, reason: "..."); optionally refuse-on-curator-down for namespaces that opted into auto_atomise_mode = Synchronous.
  4. PERF-7 fix: truncate each candidate’s content to a configurable cap (default 1500 chars / ~400 tokens) before inlining into the LLM prompt; add a synthesis_prompt_size telemetry counter.
  5. SEC-11 doc: document explicitly that synthesis is a quality gate, not a security gate; add per-namespace synthesis_failure_mode = "fall_through" | "block_write" knob.
  6. Tests: synthesis_delete_verdict_consults_k9_per_candidate, synthesis_response_carries_synthesis_failed_on_llm_error, synthesis_prompt_truncates_candidate_content_at_cap.
• Files:
  • /Users/fate/v07/synth-fix/src/synthesis/mod.rs
  • /Users/fate/v07/synth-fix/src/mcp/tools/store.rs
  • /Users/fate/v07/synth-fix/src/governance/ (K9 hook call site)
  • /Users/fate/v07/synth-fix/tests/form_1_synthesis.rs
• Findings absorbed: SEC-1 HIGH, SEC-11 MED, COR-5 MED, COR-6 MED, PERF-7 HIGH
• LOE: ~1 session
• Parallel-safe with: A, C, D, E, G, H, I, J
• Dependencies: Touches src/mcp/tools/store.rs — coordinate with F (memory_store hot-path refactor) by landing B first or F first, then rebase.

CLUSTER C — Signed-events chain integrity (BEGIN IMMEDIATE + drainer DLQ + HMAC negative tests)

• Scope:
  1. SEC-3 fix: replace conn.unchecked_transaction() (DEFERRED default) with conn.transaction_with_behavior(TransactionBehavior::Immediate) in append_signed_event. Audit signed_events.rs for other unchecked_transaction sites.
  2. Drainer in deferred_audit.rs requeues on SQLITE_CONSTRAINT_UNIQUE specifically; add a DLQ for unrecoverable failures (counter-exposed via capabilities).
  3. COR-9 fix: read_chain_head WHERE sequence IS NOT NULL (don’t COALESCE NULL to 0); emit a clear migration-needed diagnostic when a NULL-sequence row is observed post-v34.
  4. COV-5 fix: add negative cross-method + cross-pending_id tests: hmac_cross_method_binding_rejected, hmac_cross_pending_id_binding_rejected.
• Files:
  • /Users/fate/v07/synth-fix/src/signed_events.rs
  • /Users/fate/v07/synth-fix/src/governance/deferred_audit.rs
  • /Users/fate/v07/synth-fix/src/atomisation/mod.rs:656 (SEC-19 site — closed by SEC-3 fix)
  • /Users/fate/v07/synth-fix/tests/signed_events_chain_v34.rs
  • /Users/fate/v07/synth-fix/tests/k10_approval_security.rs
• Findings absorbed: SEC-3 HIGH, SEC-19 LOW, COR-9 MED, COV-5 HIGH
• LOE: ~1 session
• Parallel-safe with: A, B, D, E, G, H, I, J
• Dependencies: none

CLUSTER D — L1-6 governance fail-open + handle_deref IDOR + matcher correctness

• Scope:
  1. SEC-2 fix: at daemon startup, tracing::error! (not warn) when L1-6 not active AND governance_rules is non-empty; surface l1_6_attest: false in capabilities v3; add [governance] require_operator_pubkey = true config knob for fail-closed mode.
  2. SEC-4 fix: add agent_id parameter to handle_deref (matching handle_offload), thread to ContextOffloader::deref, refuse with OffloadError::NotFound (leak-resistant) when caller != stored agent_id AND K9 does not grant cross-agent read.
  3. SEC-12 + COR-10 fix: either rename command_regex → command_substring AND add validator that rejects regex metacharacters, OR actually implement regex matching gated on a command_regex_kind: "regex" discriminator. Recommend the validator+rename for v0.7.0; regex impl for v0.7.1.
  4. SEC-13 fix: plumb actual argv into ProcessSpawn wire-point at hooks/executor.rs:469-478; extend matcher with optional args_contain / args_regex.
  5. SEC-17 fix: use as_os_str() in command_str wire-check; convert to lossy only for log surface.
  6. COV-8 fix: add named test tests/k9_kg_invalidate_governance_gate.rs::handle_kg_invalidate_refuses_when_rule_denies.
• Files:
  • /Users/fate/v07/synth-fix/src/governance/rules_store.rs
  • /Users/fate/v07/synth-fix/src/governance/agent_action.rs
  • /Users/fate/v07/synth-fix/src/hooks/executor.rs
  • /Users/fate/v07/synth-fix/src/mcp/tools/offload.rs
  • /Users/fate/v07/synth-fix/src/offload/mod.rs
  • /Users/fate/v07/synth-fix/src/capabilities.rs
• Findings absorbed: SEC-2 HIGH, SEC-4 HIGH, SEC-12 MED, SEC-13 MED, SEC-17 LOW, COR-10 MED, COV-8 MED
• LOE: ~1 session
• Parallel-safe with: A, B, C, E, G, H, I, J
• Dependencies: none

CLUSTER E — Form 6 kind-filter inversion + L1-5 skill HTTP/CLI parity + capability skills CLI

• Scope:
  1. COR-4 fix: distinguish parsed-zero from omitted in MemoryKind::parse_csv. Return Some(vec![]) for empty-intentional → no matches; None for absent → no filter. Update apply_kinds_filter. Cascading test fixes.
  2. API-3 fix: add #[arg(long = "kinds", alias = "kind")] on CLI; update doc examples.
  3. Skills parity slice (subset of API-2): ship ai-memory skill {register|list|get|export|promote|compose} CLI subcommands AND POST /api/v1/skill/{register,list,get,export,promote,compose} HTTP routes. The handlers already exist in src/mcp/tools/skill_*.rs; the work is boilerplate copy-paste into the CLI dispatcher + Axum router.
  4. Tests: mcp_handler_recall_kinds_empty_array_returns_zero_results, cli_recall_kinds_alias_kind_still_works, skill CLI/HTTP integration tests (6 each).
• Files:
  • /Users/fate/v07/synth-fix/src/models/memory.rs (parse_csv)
  • /Users/fate/v07/synth-fix/src/mcp/tools/recall.rs (apply_kinds_filter)
  • /Users/fate/v07/synth-fix/src/cli/recall.rs
  • /Users/fate/v07/synth-fix/src/cli/skill.rs (NEW)
  • /Users/fate/v07/synth-fix/src/handlers/http.rs (NEW skill routes)
  • /Users/fate/v07/synth-fix/src/lib.rs (router additions)
  • /Users/fate/v07/synth-fix/tests/form_6_memorykind_vocab.rs
  • /Users/fate/v07/synth-fix/tests/skill_test.rs
• Findings absorbed: COR-4 HIGH, API-3 MED, API-2 HIGH (Skills slice only — other HTTP gaps in cluster K disposition).
• LOE: ~1.5 sessions (skills CLI+HTTP is the bulk)
• Parallel-safe with: A, B, C, D, G, H, I, J
• Dependencies: none

CLUSTER F — Performance hot-paths (memory_store / memory_recall)

• Scope:
  1. PERF-1 fix: thread &Connection from MCP handler into maybe_enqueue_auto_atomise, auto_persona, auto_export rather than calling db::open per hook. Removes 5+ SQLite syscalls per store.
  2. PERF-2 fix: in recall hybrid loop, extract embedding bytes in the row mapper (already in the SELECT); move cosine calc into row-iteration body. Eliminates N+1 get_embedding queries.
  3. PERF-6 fix: batch all K touches into a single BEGIN IMMEDIATE transaction; merge the 3 UPDATEs into one CASE-driven statement.
  4. PERF-10 fix: change maybe_enqueue_auto_atomise / run_synchronous_auto_atomise to accept &Memory + &str id; remove ..mem.clone() spread.
  5. PERF-14 fix: swap Vec<Memory> clone for &[&Memory] in synthesise.
  6. PERF-15 fix: subordinate to PERF-1.
  7. PERF-5 fix: treat Form-2 Synchronous mode as operator contract — document latency envelope in auto_atomise_mode docstring; add per-namespace deadline fall-through to Deferred on timeout; consider reducing default max_retries to 1.
  8. Tests: regression test asserting recall no longer issues separate get_embedding; regression test asserting K touches share one writer transaction.
• Files:
  • /Users/fate/v07/synth-fix/src/mcp/tools/store.rs (handler signature + hook wire-up)
  • /Users/fate/v07/synth-fix/src/hooks/pre_store/auto_atomise.rs
  • /Users/fate/v07/synth-fix/src/hooks/post_reflect/auto_persona.rs
  • /Users/fate/v07/synth-fix/src/hooks/post_reflect/auto_export.rs
  • /Users/fate/v07/synth-fix/src/storage/mod.rs (recall hybrid + touch)
  • /Users/fate/v07/synth-fix/src/synthesis/mod.rs
  • /Users/fate/v07/synth-fix/src/atomisation/curator.rs (max_retries)
• Findings absorbed: PERF-1 CRIT, PERF-2 CRIT, PERF-5 HIGH, PERF-6 HIGH, PERF-10 MED, PERF-14 LOW, PERF-15 LOW
• LOE: ~2 sessions (handler signature change cascades; well-scoped)
• Parallel-safe with: C, G, H, I, J, K, L
• Dependencies: Sequential after B (Cluster B touches src/mcp/tools/store.rs synthesis call site — must land first or merge-conflict). After F, all of A/B/D have touched the same area, so coordinate via the Group-1/Group-2 schedule below.

CLUSTER G — Form 5 shadow-mode unboundedness + sampling + calibration streaming

• Scope:
  1. PERF-4 fix: add shadow_retention_days config (default 30) + sweeper in existing background GC scheduler; add compound (namespace, source, observed_at) index; stream the calibration aggregation (in-SQL GROUP BY or chunked scroll cursor).
  2. PERF-9 fix: read AI_MEMORY_CONFIDENCE_SHADOW + _SAMPLE_RATE env vars once at daemon boot into a OnceLock<ShadowConfig>; gate observe() on cached value.
  3. PERF-12 fix: rewrite calibration sweep as SELECT namespace, source, COUNT(*), AVG(...), … aggregate; use streaming median (Welford-style) over sorted cursor when percentile_cont unavailable.
  4. COV-2 fix: add mcp_handler_calibrate_confidence_returns_baselines_envelope test.
  5. COV-14 fix: add recall_touch_with_decay_env_set_updates_decayed_at test.
• Files:
  • /Users/fate/v07/synth-fix/migrations/sqlite/0033_v07_form5_confidence_calibration.sql (or new follow-up migration for the compound index + retention column)
  • /Users/fate/v07/synth-fix/src/confidence/shadow.rs
  • /Users/fate/v07/synth-fix/src/confidence/calibrate.rs
  • /Users/fate/v07/synth-fix/src/background/ (new sweeper)
  • /Users/fate/v07/synth-fix/tests/form_5_confidence_calibration.rs
• Findings absorbed: PERF-4 HIGH, PERF-9 MED, PERF-12 MED, COV-2 HIGH, COV-14 LOW
• LOE: ~1.5 sessions (GC sweeper + retention column = bulk)
• Parallel-safe with: A, B, C, D, E, H, I, J, K, L
• Dependencies: none

CLUSTER H — Docs + cookbook + CLI-help accuracy sweep (DOCS-ONLY)

• Scope:
  1. DOC-4 / API-1 / API-7 / API-20 (tool count drift): single source of truth. build.rs substitution OR CI grep gate against Profile::full().expected_tool_count(). Update README badge 7_default • 71_full, narrative “7 always-loaded … 71 total”, release-notes “60 → 71”, daemon_runtime.rs:181 --profile full help string from “43 tools” to “71 tools at v0.7.0”.
  2. DOC-2 (release notes): append Post-grand-slam Form 1-6 + 7th-form + QW wave section to docs/v0.7.0/release-notes.md mirroring docs/internal/v070-feature-inventory.md.
  3. DOC-3 (README “5 substrates”): rewrite “What’s new in v0.7” anchored on canonical feature inventory; link docs/atomisation.md, docs/persona.md, docs/memory-kind-vocab.md, docs/confidence-calibration.md, docs/provenance.md.
  4. DOC-5: flip evidence-frozen badge from v0.6.4 to v0.7.0 (or add a v0.7.0 companion badge).
  5. DOC-6/7/8/10/11: rewrite docs/MIGRATION_v0.7.md — delete TODO — track X lands blocks; rename memory_approval_pending/memory_approval_decide → memory_pending_list/memory_pending_approve/memory_pending_reject; fix “20 lifecycle events” → “25”; either drop doctor --kg-backend claim or ship the flag (recommend drop the claim, document ai-memory doctor --remote JSON grep); add per-form sections.
  6. DOC-9: in docs/memory-kind-vocab.md, replace ai-memory doctor --capabilities=v3 with ai-memory mcp call memory_capabilities '{"schema_version":"3"}' | jq .memory_kind_vocab.
  7. DOC-12: in cookbook/multistep-ingest/01-two-phase.sh, replace hardcoded /Users/fate/v07/v07-fixes/ paths with "${REPO_ROOT}/.local-runs" / "${REPO_ROOT}/target".
  8. DOC-13: add post-audit prologue to docs/internal/batman-framework-audit.md declaring it reflects state at 53b4d39 and Forms 1-6 + 7th shipped in response (PRs #761-#766).
  9. DOC-14/15: sweep new HTTP routes from src/handlers/http.rs into docs/API_REFERENCE.md; add §”v0.7.0 new commands” block to docs/CLI_REFERENCE.md.
  10. DOC-16: ship MVP docs for Hook pipeline, Federation hardening, K8 quotas, K10 SSE approvals, Sidechain transcripts, Signed-events V-4 chain. 200-500 lines each in docs/atomisation.md style.
  11. DOC-17: create RELEASE_NOTES_v0.7.0.md at project root (synopsis + pointer to docs/v0.7.0/release-notes.md) OR document convention in CONTRIBUTING.md.
  12. DOC-18: verify Gemma model identifier (gemma3:4b vs “Gemma 4 E2B”) between docs/atomisation.md and entrypoint.plan-c.sh; reconcile.
  13. DOC-19: add reconcile banner to docs/policy-engine.md.
  14. API-5 (error-code convention ADR): define ADR — recommend lowercase.dotted as canonical; add alias map; update tests.
  15. API-6: add /api/v1/load_family alias for /api/v1/memory_load_family with Deprecation header; mention in MIGRATION_v0.7.md.
  16. API-8/11: rename CLI envelope pending_id → id (or add alias); replace --from-shadow bool with --source <SOURCE> enum.
  17. API-10: remove HTTP-admin-endpoints claim from src/mcp/registry.rs:1325 docstring.
  18. API-4: add "source": {"enum": ["shadow"], "default": "shadow"} to MCP memory_calibrate_confidence inputSchema.
• Files: (many; docs and registry strings)
  • /Users/fate/v07/synth-fix/README.md
  • /Users/fate/v07/synth-fix/docs/v0.7.0/release-notes.md
  • /Users/fate/v07/synth-fix/docs/MIGRATION_v0.7.md
  • /Users/fate/v07/synth-fix/docs/memory-kind-vocab.md
  • /Users/fate/v07/synth-fix/docs/atomisation.md
  • /Users/fate/v07/synth-fix/docs/policy-engine.md
  • /Users/fate/v07/synth-fix/docs/internal/batman-framework-audit.md
  • /Users/fate/v07/synth-fix/docs/API_REFERENCE.md
  • /Users/fate/v07/synth-fix/docs/CLI_REFERENCE.md
  • /Users/fate/v07/synth-fix/cookbook/multistep-ingest/01-two-phase.sh
  • /Users/fate/v07/synth-fix/src/daemon_runtime.rs:181 (help string)
  • /Users/fate/v07/synth-fix/src/mcp/registry.rs:1325 (docstring)
  • /Users/fate/v07/synth-fix/src/cli/governance.rs (envelope key rename)
  • /Users/fate/v07/synth-fix/src/cli/commands/calibrate_confidence.rs (–source enum)
  • NEW: docs/hooks.md, docs/federation.md, docs/quotas.md, docs/approvals.md, docs/transcripts.md, docs/signed-events.md
  • NEW: RELEASE_NOTES_v0.7.0.md (root)
• Findings absorbed: DOC-2 CRIT (DOC-1 RESOLVED separately), DOC-3/4/5/6/7/8/9/10/11/12/13/14/15/16/17/18/19 (all), API-1 HIGH, API-2 HIGH (docs slice — full slice in K), API-4 MED, API-5 MED, API-6 MED, API-7 MED, API-8 LOW, API-10 INFO, API-11 LOW, API-20 LOW
• LOE: ~3 sessions (DOC-16 alone is 12-20 hours of net-new doc writing; split into a sub-cluster H-2 if needed)
• Parallel-safe with: all code clusters (docs-only — minimal overlap with src/; only three Rust touchpoints: daemon_runtime.rs, mcp/registry.rs docstring, cli/commands/calibrate_confidence.rs)
• Dependencies: none

CLUSTER I — CI test-runner expansion (postgres integration tests not running)

• Scope:
  1. COV-1 fix: change .github/workflows/ci.yml lines 211, 252 from cargo test --features sal-postgres --lib and cargo test --features sal --lib to cargo test --features sal-postgres --workspace (or drop --lib). Verify the pgvector/pgvector:pg16 service container + AI_MEMORY_TEST_POSTGRES_URL env var are wired through.
  2. COV-6 fix: add tests/postgres_schema_parity.rs::pre_form6_postgres_db_upgrades_memory_kind_column_idempotently.
  3. COV-7 fix: add tests/schema_ladder_v14_to_v39.rs::legacy_v064_db_forwards_migrates_through_all_arms with a checked-in v0.6.4 seed DB (~50 KB) generated via git checkout v0.6.4 && cargo run -- init && git checkout -.
  4. COV-9 fix: tests/offload/ttl_sweep_loop.rs::spawn_loop_drains_expired_within_interval_window.
  5. COV-10 fix: mcp_handler_dependents_envelope_round_trip.
  6. COV-11 fix: tests/persona/generate_acceptance.rs::generate_distils_cluster_into_persona_with_citations with a MockLlmDispatch.
  7. COV-12 fix: pre_v070_peer_payload_deserialises_with_default_provenance.
  8. COV-13 fix: mcp_handler_deref_envelope_round_trip + mcp_handler_deref_refuses_tampered_blob_with_structured_error.
  9. COV-4 fix: auto_persona_hook_fires_every_n_memories end-to-end test.
• Files:
  • /Users/fate/v07/synth-fix/.github/workflows/ci.yml
  • /Users/fate/v07/synth-fix/tests/ (numerous new files)
  • Seed DB: /Users/fate/v07/synth-fix/tests/fixtures/v064_seed.db (or equivalent)
• Findings absorbed: COV-1 CRIT, COV-4 HIGH, COV-6 HIGH, COV-7 MED, COV-9 MED, COV-10 MED, COV-11 MED, COV-12 MED, COV-13 MED
• LOE: ~2 sessions (CI gate flip = 0.5h; the 8 test adds = remainder)
• Parallel-safe with: all (test-only)
• Dependencies: Sequential after A (COV-3 belongs in A; once A lands, COV-3 closes organically; if A is delayed, COV-3 stays open here). The other adds are independent.

CLUSTER J — Migration filename collisions cleanup

• Scope:
  1. COR-13 / DOC-20 fix: rename migrations/sqlite/0031_v07_namespace_auto_atomise.sql → migrations/sqlite/0031_v07_namespace_auto_atomise.NOTES.md (or move under docs/internal/migration-notes/). Same for migrations/postgres/0018_v07_persona.sql vs 0018_v07_namespace_auto_atomise.sql.
  2. Verify grep -rn "0031_v07_namespace_auto_atomise" src/ returns zero matches (then safe to rename — currently it returns zero per DOC-20).
  3. Add tests/migration_filename_unique.rs::no_duplicate_migration_prefix_collisions ship-gate.
• Files:
  • /Users/fate/v07/synth-fix/migrations/sqlite/0031_v07_namespace_auto_atomise.sql
  • /Users/fate/v07/synth-fix/migrations/sqlite/0031_v07_persona.sql (verify slot)
  • /Users/fate/v07/synth-fix/migrations/postgres/0018_v07_namespace_auto_atomise.sql
  • /Users/fate/v07/synth-fix/migrations/postgres/0018_v07_persona.sql
  • NEW: /Users/fate/v07/synth-fix/tests/migration_filename_unique.rs
• Findings absorbed: COR-13 LOW, DOC-20 LOW
• LOE: 30 minutes
• Parallel-safe with: all
• Dependencies: none

CLUSTER K — QW-4 honest disposition (operator decision)

• Scope:
  • Decision point — EITHER: (a) Ship QW-4 (Tencent’s “fourth quick win”; competitive positioning beyond docs). Scope, design, implement, test in v0.7.0 — would push tag-cut. (b) Re-attribute QW-4 to Form 7 governance refusals (MCP_MUTATION_DISABLED_ERROR is a plausible QW-4 anchor) and update CHANGELOG + release-notes accordingly. (c) Defensibly defer to v0.7.1 — remove QW-4 from “shipped” lists in CHANGELOG / release-notes / docs/positioning.md; document as “scoped for v0.7.1”.
  • Recommendation: (c) defer — fastest path to tag-cut with no procurement lie.
  • HTTP-coverage gap (API-2 minus the Skills slice handled in E): same operator decision. 13 missing routes; recommend (b) “honest documentation” in MIGRATION_v0.7.md noting the MCP-only surface, with HTTP routes scoped for v0.7.1.
• Files: (depending on decision)
  • /Users/fate/v07/synth-fix/CHANGELOG.md
  • /Users/fate/v07/synth-fix/docs/v0.7.0/release-notes.md
  • /Users/fate/v07/synth-fix/docs/positioning.md
  • /Users/fate/v07/synth-fix/docs/internal/v070-feature-inventory.md (KU #9 closure note)
• Findings absorbed: API-12 HIGH, API-2 HIGH (residual after E), DOC implicit
• LOE: Option (a) 1-2 sessions; (b) 30 minutes; (c) 30 minutes
• Parallel-safe with: all
• Dependencies: OPERATOR DECISION REQUIRED

CLUSTER L — INFO / accepted-debt / operator-document-only (no fix-agent dispatch)

These are real findings but operator-defensible or explicitly scoped out of v0.7.0. Flag for documentation rather than code change.

• SEC-5 (case-sensitive host: filter) — LOW. Defense-in-depth; lowercase the compare in two sites. Could roll into D as a 30-min add-on (recommend).
• SEC-6 (single-key HTTP auth) — MEDIUM but documented as design (out of scope for v0.7.0 per the federation hardening doc). Document the limitation in security posture page.
• SEC-7 (federation x-peer-id is operator-bound not cert-bound) — MEDIUM. v0.8.0 closeout tracked; v0.7.0 mitigation = startup-time WARN on bypass env vars (1 hour; recommend roll into D).
• SEC-8 (validate_id accepts ..///\) — MEDIUM. Tighten validator + add \ to filesystem-emit sanitisers (recommend roll into D as add-on).
• SEC-9 (SSRF validate_url_dns fails open on DNS hiccup) — MEDIUM, documented design. Optional validate_at_dispatch_too opt-in for hardening; recommend defer.
• SEC-10 (quota_status K9-rule permissive enumeration) — MEDIUM, docs-only. Roll into H.
• SEC-14 (validate_namespace “..” only at segment level) — LOW. Defense-in-depth.
• SEC-15 (auto-export detached-thread silent failure) — LOW. Counter add (recommend defer).
• SEC-16 (persona entity_id minimal validation) — LOW. Roll into D.
• SEC-18 (env-prefix redactor case sensitivity) — LOW. Add pass / pwd to keyword list.
• SEC-20 (K10 SSE 1024 channel capacity, no per-agent rate-limit) — INFO, v0.8.0 scope.
• SEC-21 (publish-sdks.yml tag-trigger) — INFO, procurement-grade acceptable.
• SEC-22 (migration 0025 backfill) — INFO, safe.
• SEC-23 (Form 5 shadow honors contract) — INFO, positive.
• COR-11 (last_err unreachable placeholder) — LOW, cosmetic.
• COR-12 (env-var test races) — LOW. serial_test crate or process mutex. Roll into G.
• COR-8 (archive_source BEGIN IMMEDIATE in outer tx) — MEDIUM. Use unchecked_transaction + nested-tx detection. Recommend roll into A.
• PERF-8 (auto_persona LIKE %X% scan) — MEDIUM. Requires schema column extension + backfill; recommend defer (v0.7.1 climb-back).
• PERF-11 (Form 3 content duplication across stages) — MEDIUM. Recommend defer.
• PERF-13 (deferred_audit unbounded channel) — MEDIUM. Recommend roll into C (signed_events cluster).
• PERF-16 (format! in Form 1 candidate loop) — LOW, 5-iter bound. Defer.
• PERF-17 (auto_persona resolve_entity_id JSON parse) — LOW. Subordinate to PERF-8.
• COV-15/16/17/18 — LOW, opportunistic test adds. Recommend defer.
• API-9 / API-13/14/15/16/17/18/19 — INFO. Net-zero confirmations.

Already-fixed (verified at 0536e96)

• DOC-1 CHANGELOG silent for Forms/QW/WT-1 → RESOLVED by commit bbeb1e8 (docs(CHANGELOG): comprehensive v0.7.0 entry update — WT-1 + QW + Batman 6+7 + security reconciliation). Verified via grep -c "Form|atomis|persona|offload|calibrate" CHANGELOG.md → 59 matches.
• Postgres init retry loop (PERF “areas requiring investigation” #4) → addressed by 0536e96 (fix(v0.7.0 ship-readiness): postgres concurrent-init retry + g3 timeout + stderr-redaction test alignment). Reviewer-Performance could not locate this in the baseline; it landed AFTER the reviewer’s anchor.
• g3 timeout flake (Flake census #1) → addressed by 0536e96 (same commit).
• stderr-redaction test alignment → addressed by 0536e96.

Parallel-safe dispatch order

Group 1 (parallel, low/no shared-write-surface overlap):

• CLUSTER A (Form 4 code paths — atomisation/, storage/, validate.rs)
• CLUSTER C (signed_events.rs + deferred_audit.rs)
• CLUSTER D (governance/ + hooks/executor.rs + mcp/tools/offload.rs)
• CLUSTER E (MemoryKind enum + Skills CLI/HTTP — new files mostly)
• CLUSTER G (confidence/shadow.rs + calibrate.rs + new background sweeper + new migration)
• CLUSTER H (docs sweep — touches src/ in only 3 small spots)
• CLUSTER I (CI workflow + new test files)
• CLUSTER J (migration filename rename + new test)

Group 2 (after Group 1; touch overlapping write-path files):

• CLUSTER B (Form 1 synthesis — src/mcp/tools/store.rs synthesis call site + src/synthesis/mod.rs)
• CLUSTER F (memory_store + memory_recall hot-path refactor — src/mcp/tools/store.rs handler signature + src/storage/mod.rs recall path)

Coordination notes:

• B and F both touch src/mcp/tools/store.rs. Land B first (smaller, security-load-bearing), then F (signature refactor that can rebase cleanly).
• A and F both touch src/storage/mod.rs but in different functions (row_to_memory vs recall_hybrid/touch); should not conflict.
• D and B both consult K9 governance — D lands the wire-points; B consumes them. Land D first if you want B’s K9-on-delete to use the cleaned-up matcher; otherwise either order works.

Group 3 (operator decision):

• CLUSTER K (QW-4 disposition + HTTP-coverage policy decision)

Already-fixed verification:

• Confirm bbeb1e8 actually covers DOC-1 (done — 59 matches).
• Confirm 0536e96 covers the postgres init retry / g3 timeout / stderr-test alignment (per commit message — done).

Operator decision points (cannot autonomously resolve)

1. CLUSTER K disposition — QW-4: ship (a), re-attribute (b), or defer (c)? Recommendation: (c) defer, accompanied by HTTP-coverage gap honesty-doc in MIGRATION_v0.7.md.
2. CLUSTER H sub-scoping — DOC-16 alone is 12-20 hours. Split off as cluster H-2 “Hook pipeline + Federation + K8 quotas + K10 SSE + Transcripts + Signed-events dedicated docs” and defer to v0.7.0.1 doc-only point release? Or block tag-cut on it?
3. CLUSTER E Skills HTTP routes — confirm CLI Skills are wanted alongside HTTP (operator may prefer MCP-only by design); if so, downscope E to MemoryKind+CLI alias only.
4. CLUSTER F PERF-5 default — reduce max_retries from 3 to 1 on Synchronous path? This is a behavior change visible to operators relying on aggressive retry.
5. API-5 error-code convention ADR — pick UPPER_SNAKE vs lowercase.dotted as canonical (recommend lowercase.dotted).
6. LOE > 2-session clusters:
   • H (docs sweep, ~3 sessions; split as H-1 stale-fix and H-2 net-new MVP docs).
   • F if PERF-5 default change requires a deprecation cycle.
7. API-6 breaking-change alias decision — keep /api/v1/memory_load_family working with Deprecation header (recommend) vs hard-rename (breaks v0.7.0-rc users).

End-of-synthesis checklist

• All 6 reviewer deliverables read in full at their pinned commits.
• Verified DOC-1 (CHANGELOG silent) RESOLVED by bbeb1e8 (59 matches for Form/atomise keywords in current CHANGELOG.md at 0536e96).
• Counted raw findings: 23 + 13 + 17 + 20 + 20 + 18 = 111 (prompt said 103; delta noted in §”Reviewer source counts”).
• Dedup applied per “Known dedupes” table.
• 12 clusters defined with file paths, findings absorbed, LOE, parallel-safety, deps.
• Group 1 / Group 2 dispatch schedule provided.
• Operator decision points enumerated (7).
• Already-fixed items distinguished from open work.

— Cold mountain.