docs: restore trimmed design record after audit review

Fix stray fence in architecture.md; mark Step 1 open questions resolved;
restore apply-engine citations, Q6 SARIF de-dup shape, and profiling note.

Author: SutuSebastian

docs/plans/github-marketplace-action.md

@@ -133,7 +133,7 @@ These are the design questions the plan-PR resolves before impl starts. Each get
 
   **Resolution (Q-B grill):** **α + skip-on-non-PR**. On `pull_request` events the Action runs `codemap audit --base ${{ github.base_ref }} --ci` (SARIF + non-zero exit + quiet). On any other event (`push`, `schedule`, `workflow_dispatch`, …) the Action no-ops with `echo "codemap action: no PR context, skipping"` and exits 0 — there's no meaningful default without a base ref, and firing β-style opinionated recipes on push events surfaces findings consumers didn't ask for. β is a worse α (no PR-scoped diff). γ is reserved until multi-recipe SARIF composition is proven; current `action.yml` rejects `mode: aggregate` with a v1.x message. δ over-engineers a config concept that doesn't exist. Consumers wanting push-event runs pass an explicit `command:` input.
 
-- **Q6 — SARIF rule.id taxonomy under `mode: aggregate`.** _(Deferred to v1.x — superseded Q1 ripple that promoted this to a v1.0 blocker.)_ Single-recipe `--format sarif` is shipped (`rule.id = codemap.<recipe-id>`). Aggregate runs (multiple recipes in one Action invocation) need a stable convention for combining results — likely concat-into-one-sarif-file with rule definitions de-duplicated by `recipe-id`. Verify in GitHub Code Scanning before enabling `mode: aggregate`; fall back to one-SARIF-per-recipe with `category:` if dedup misbehaves.
+- **Q6 — SARIF rule.id taxonomy under `mode: aggregate`.** _(Deferred to v1.x — superseded Q1 ripple that promoted this to a v1.0 blocker.)_ Single-recipe `--format sarif` is shipped (`rule.id = codemap.<recipe-id>`). Aggregate runs (multiple recipes in one Action invocation) need a stable convention for combining results — likely concat-into-one-sarif-file with rule definitions de-duplicated by `recipe-id` (each rule appears once in `tool.driver.rules[]`; each finding's `ruleId` matches). Verify in GitHub Code Scanning before enabling `mode: aggregate`; fall back to one-SARIF-per-recipe with `category:` if dedup misbehaves.
 - **Q7 — Versioning + `@v1` tag strategy.** Convention: floating `@v1` major tag updated on every minor/patch release; `@v1.2.3` for pin-to-exact. `action.yml` lives at repo root; `dist/` is **not** required for composite actions (only for JS actions).
 
   **Resolution: floating major `@v1` + changesets-driven release.** Cohort norm (`actions/checkout@v4`, `actions/setup-node@v4`, `oven-sh/setup-bun@v2`, …); Renovate/Dependabot-friendly (major tags are the unit they bump on); enforces semver discipline (any input/output break is a v2 bump, never within v1.x).

docs/plans/perf-triangulation-rollout.md

@@ -84,7 +84,7 @@ Every distinct finding across the five source audits, with per-audit priority. E
 
 The audit's predicted optimisation (per-file `Map.get` hoist saving the ~38ms it estimated from ~1.26M skipped lookups × ~30ns each) was **wrong** — tested both `ORDER BY file_path, id` SQL and JS-side `Map<file, Ref[]>` grouping variants on this repo (340 files) and a 2.1k-file external corpus; both showed 0 to slight regression. V8 already optimises hot `Map.get`; JS-side grouping overhead exceeded any savings.
 
-**Profile-driven actual win:** Bindings-phase profiling revealed `bindings_ms` decomposes as `resolveBindings ≈ 17%` + `persistBindings.insert ≈ 83%` on a 2k-file corpus — the bottleneck was 243k row INSERTs with `foreign_keys=ON` + `synchronous=NORMAL` per row, NOT the resolver loop. Extending the existing bulk-INSERT PRAGMA-OFF window (already used during parallel parse+insert) through the bindings/cycles/re-exports phase saved **-33% `bindings_ms`** on the 2k-file corpus and **-27% here**. Behavior-preserved (stable-snapshot SHA bit-identical on both corpora).
+**Profile-driven actual win:** Bindings-phase profiling (`bindings_ms` decomposition instrumentation; no longer a public env var) revealed `bindings_ms` decomposes as `resolveBindings ≈ 17%` + `persistBindings.insert ≈ 83%` on a 2k-file corpus — the bottleneck was 243k row INSERTs with `foreign_keys=ON` + `synchronous=NORMAL` per row, NOT the resolver loop. Extending the existing bulk-INSERT PRAGMA-OFF window (already used during parallel parse+insert) through the bindings/cycles/re-exports phase saved **-33% `bindings_ms`** on the 2k-file corpus and **-27% here**. Behavior-preserved (stable-snapshot SHA bit-identical on both corpora).
 
 The "deeper optimisations (TypedArrays, no-imports fast-path)" the original audit gated on a larger corpus are still untested and may be similarly off — see § Methodology gaps below.
 

docs/research/codemap-richer-index-synthesis-2026-05.md

@@ -71,16 +71,16 @@ A1 and A4 share an author but diverge on the agent-angle weighting and AST-engin
 
 These have **6/6** support (counting A1 and A4 separately because their conclusions diverge on emphasis even when the author overlaps). They are the load-bearing claims for the synthesised path; nothing in the disagreement map should override them.
 
-| #   | Claim                                                                                                                                                                                                                                                                                       | Supported by                                     |
-| --- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------ |
-| C1  | The doc reframe (**shipped 2026-05**): [`roadmap.md`](../roadmap.md) Floors now split verdict-shaped lints from the substrate-shaped `codemap apply` executor. The old single "no fix engine" row conflated two product classes.                                                            | A1, A2, A3, A4, A5, A6                           |
-| C2  | The reframe shape: **"No verdict-shaped fix engine"** (preserved — `knip` / `eslint --fix` / `jscpd` own that class) AND **"Substrate-shaped diff executor IS in scope"** (the shipped reality; recipes propose, engine executes, agent confirms).                                          | A1, A2 (implicit), A3 (implicit), A4, A5, A6     |
-| C3  | `calls.{line_start, column_start}` is the **highest-ROI single substrate addition.** Closes the rename-preview substrate gap; unblocks app-wide rename of call sites; foundation for replace-deprecated-call recipes; foundation for diagnostic-push hotspot rendering when LSP plan ships. | A1, A2, A3, A4, A5, A6                           |
-| C4  | **Re-export source locations** are needed (whether as `exports.{line_start, column_start}` columns or a separate `export_locations` table). Re-export alias chains and barrel-safe rewrites both need it.                                                                                   | A1, A2, A3, A4, A5, A6                           |
-| C5  | More diff-shape recipes are needed in the bundled catalog. Today the count is 1 (`rename-preview`); consensus floor is 3–5 more. Specific candidates with multi-source agreement listed in [§ 4.3](#43-recipe-additions).                                                                   | A1, A2, A3, A4 (item #4), A5, A6 (implicit)      |
-| C6  | **Moat A applies to writes by analogue.** Reviewer test: _"is this fix also expressible as `query --recipe <id>` + `apply <recipe-id>`?"_ Pass → moat-clean. Fail → verdict shape; reject. The recipes are the API for writes, same as for reads.                                           | A1, A2, A3, A4, A5, A6                           |
-| C7  | **Dry-run-first + explicit confirmation gate is correct** and stays. Phase-1-validates-before-phase-2-writes pattern preserved; non-TTY without `--yes` rejected.                                                                                                                           | A1, A2, A3, A4, A5, A6                           |
-| C8  | **Watch + apply + reindex is the agent loop.** Default-ON watch on long-running transports keeps reads fresh; codemap-specific structural verification (the index reflects disk truth) is sufficient for the agent path; full typecheck/lint/tests verifier is consumer-side.               | A1, A4, A6 explicit; A2, A3, A5 don't contradict |
+| #   | Claim                                                                                                                                                                                                                                                                                                                                                         | Supported by                                     |
+| --- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------ |
+| C1  | The doc reframe (**shipped 2026-05**): [`roadmap.md`](../roadmap.md) Floors now split verdict-shaped lints from the substrate-shaped `codemap apply` executor (`apply-engine.ts`, `cmd-apply.ts`, `rename-preview.sql`; see [`glossary.md`](../glossary.md) Substrate-shaped fix executor). The old single "no fix engine" row conflated two product classes. | A1, A2, A3, A4, A5, A6                           |
+| C2  | The reframe shape: **"No verdict-shaped fix engine"** (preserved — `knip` / `eslint --fix` / `jscpd` own that class) AND **"Substrate-shaped diff executor IS in scope"** (the shipped reality; recipes propose, engine executes, agent confirms).                                                                                                            | A1, A2 (implicit), A3 (implicit), A4, A5, A6     |
+| C3  | `calls.{line_start, column_start}` is the **highest-ROI single substrate addition.** Closes the rename-preview substrate gap; unblocks app-wide rename of call sites; foundation for replace-deprecated-call recipes; foundation for diagnostic-push hotspot rendering when LSP plan ships.                                                                   | A1, A2, A3, A4, A5, A6                           |
+| C4  | **Re-export source locations** are needed (whether as `exports.{line_start, column_start}` columns or a separate `export_locations` table). Re-export alias chains and barrel-safe rewrites both need it.                                                                                                                                                     | A1, A2, A3, A4, A5, A6                           |
+| C5  | More diff-shape recipes are needed in the bundled catalog. Today the count is 1 (`rename-preview`); consensus floor is 3–5 more. Specific candidates with multi-source agreement listed in [§ 4.3](#43-recipe-additions).                                                                                                                                     | A1, A2, A3, A4 (item #4), A5, A6 (implicit)      |
+| C6  | **Moat A applies to writes by analogue.** Reviewer test: _"is this fix also expressible as `query --recipe <id>` + `apply <recipe-id>`?"_ Pass → moat-clean. Fail → verdict shape; reject. The recipes are the API for writes, same as for reads.                                                                                                             | A1, A2, A3, A4, A5, A6                           |
+| C7  | **Dry-run-first + explicit confirmation gate is correct** and stays. Phase-1-validates-before-phase-2-writes pattern preserved; non-TTY without `--yes` rejected.                                                                                                                                                                                             | A1, A2, A3, A4, A5, A6                           |
+| C8  | **Watch + apply + reindex is the agent loop.** Default-ON watch on long-running transports keeps reads fresh; codemap-specific structural verification (the index reflects disk truth) is sufficient for the agent path; full typecheck/lint/tests verifier is consumer-side.                                                                                 | A1, A4, A6 explicit; A2, A3, A5 don't contradict |
 
 These eight are the **non-negotiable spine** of any synthesis path. Every step in [§ 6](#6-synthesised-12-step-path-with-open-implementation-questions) preserves them.
 
@@ -527,10 +527,10 @@ The minimum synthesis preserving every consensus claim (§ 2) and resolving ever
 
 **Source ancestry:** C1, C2; all six positions.
 
-**Open questions:**
+**Open questions (resolved on ship):**
 
-1. Two distinct rows or one row with two clauses? Bias toward two distinct rows — they're independent product-class decisions and each should grep / cite cleanly.
-2. Lift any salvageable text from the existing row 45 ("Adjacent to Moat A" framing)? Yes — both new rows reference Moat A.
+1. Two distinct rows or one row with two clauses? **Shipped: two distinct rows** in [`roadmap.md`](../roadmap.md) Floors.
+2. Lift any salvageable text from the existing row 45 ("Adjacent to Moat A" framing)? **Yes** — both new rows reference Moat A.
 
 ### Step 2 — Three more diff-shape recipes (S × 3)