CodeWithJuber · CodeWithJuber · Sep 26, 2026 · Sep 26, 2026 · Sep 26, 2026 · Sep 26, 2026
@@ -433,8 +433,10 @@ as the `collide_check` MCP tool.
 
 **Machine-owned doc surfaces (`src/docs_render.js`, `forge docs render`).** The
 auto-maintenance layer that keeps tables and diagrams in sync with the code registries.
-Four marker-managed blocks (commands table in README, groups and MCP-tools tables in GUIDE,
-repo-map diagram in ARCHITECTURE) are regenerated from `COMMANDS`/`GROUPS`/`TOOLS`; six
+Five marker-managed blocks (commands table in README, groups and MCP-tools tables in GUIDE,
+repo-map diagram in ARCHITECTURE, and the Mintlify changelog page, rendered from
+`CHANGELOG.md` by `src/changelog_page.js` between MDX-safe JSX-comment markers) are
+regenerated from `COMMANDS`/`GROUPS`/`TOOLS`/`CHANGELOG.md`; six
 "N MCP tools" count phrases are auto-corrected; and every mermaid block across all `.md`
 and `.mdx` files receives the branded `%%{init` theme. Registry-derived blocks are CI-gated
 errors when stale; tree-derived output is advisory.
@@ -675,22 +677,22 @@ from the tree it describes.
 ```mermaid
 %%{init: {'theme':'base','themeVariables':{'primaryColor':'#201a15','primaryTextColor':'#f2ede7','primaryBorderColor':'#372c22','lineColor':'#f26430','secondaryColor':'#272019','tertiaryColor':'#171310','edgeLabelBackground':'#201a15','clusterBkg':'#171310','clusterBorder':'#4a3b2e','fontFamily':'ui-sans-serif, system-ui, sans-serif','fontSize':'14px'},'flowchart':{'curve':'basis','padding':10,'nodeSpacing':36,'rankSpacing':44}}}%%
 flowchart LR
-  test["test<br/>133 files"]
-  src["src<br/>119 files"]
+  test["test<br/>134 files"]
+  src["src<br/>120 files"]
   landing["landing<br/>61 files"]
   research["research<br/>37 files"]
   bench["bench<br/>6 files"]
   global["global<br/>5 files"]
   scripts["scripts<br/>3 files"]
   docs["docs<br/>1 file"]
   examples["examples<br/>1 file"]
-  test -- 281 --> src
+  test -- 284 --> src
   bench -- 12 --> src
   examples -- 4 --> src
+  scripts -- 3 --> src
   test -- 3 --> global
   test -- 3 --> scripts
   test -- 2 --> bench
-  scripts --> src
   src --> global
 ```
 <!-- forge:render:repo-map:end -->
@@ -6,6 +6,44 @@ to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Security
+
+- **The claim-registry table escapes backslashes, and four patterns no longer backtrack
+  quadratically on long runs (CodeQL).** `scripts/claims-status.mjs` escaped `|` in table
+  cells but not `\`, so a trailing backslash could undo an escape (`js/incomplete-sanitization`,
+  high). The model-catalog tokenizer's trailing `.0` collapse and its URL trim (`js/polynomial-redos`,
+  high, pre-existing), the workspace-glob trim and the semantic guard's edge-punctuation trim now
+  use linear scans with identical results (checked against the old regexes).
+
+### Fixed
+
+- **`forge docs check` and `forge docs render` work on a Windows checkout.** With
+  `core.autocrlf`, Markdown checks out with CRLF line endings while generated blocks render
+  with LF, so every generated block read as stale and a render wrote LF lines into a CRLF
+  file. Blocks are now compared in LF and written back in the file's own line endings.
+
+### Added
+
+- **The docs site's changelog page is generated from `CHANGELOG.md`.** It had one hand-written
+  entry from July while thirty releases shipped. `forge docs render` now writes every release,
+  plus `[Unreleased]`, as a Mintlify `<Update>` entry with each change's headline, filter tags
+  and a link to its full notes (`src/changelog_page.js`, between MDX-safe JSX-comment markers).
+  `forge docs check` fails when the page is stale, and `scripts/bump.mjs` regenerates it in the
+  release commit.
+
+### Documentation
+
+- **The Mintlify reference pages describe the current behavior** of `forge verify` (per-package
+  coverage, pre/post binding, verifier events), `forge stack` (`available` runners),
+  `forge context` (what `COMPLETE` means, `--block`), `forge reuse` and `forge ledger`
+  (lossless keys, serve-time revalidation, one vote per event, archive reasons, conflicts,
+  `--fix --dry-run`), `forge dash` (Host check, session token) and the universal router, which
+  the site did not document at all. The landing page lists all ten native targets (OpenClaw was
+  missing).
+- **Two historical `CHANGELOG.md` entries are corrected.** 1.1.2 called the project's own
+  second-machine re-run an independent replication (now marked as a dated correction), and 1.0.0
+  had lost the `\r\n` / `\n` escapes inside two code spans.
+
 ## [1.5.0] - 2026-09-26
 
 ### Changed
@@ -127,7 +165,7 @@ evidence behind it. Scripts that read the JSON output may need to adapt:
 
 ### Added
 
-- `bench/universal-router/reproduce.sh` rebuilds the shipped router prior from pinned,
+- **`bench/universal-router/reproduce.sh` rebuilds the shipped router prior** from pinned,
   sha256-checked public inputs (`sources.json`). The refit reproduces `data/router_prior.json`
   exactly: all 176 fitted values, with only `fittedAt` different (Node v22.22.2, about 7 minutes
   on 4 vCPUs). `holdout_eval.mjs` is a new seeded 150/350 held-out experiment. It is not a
@@ -146,20 +184,23 @@ evidence behind it. Scripts that read the JSON output may need to adapt:
 
 ### Documentation
 
-- A machine-readable claim/status registry (`docs/status/claims.json`, 47 claims assessed
+- **A machine-readable claim/status registry** (`docs/status/claims.json`, 47 claims assessed
   against `d2abfa6`), with a generated table in `docs/status/README.md`. `node
   scripts/claims-status.mjs --check`, now part of the CI quality gate, fails when the registry
   is invalid, the table is stale, or a `docs/cognitive-substrate/` copy has drifted from its
   `research/` source.
-- `docs/INTEGRATIONS.md`: for every supported tool, config emission, MCP registration,
-  automatic hooks and enforcement are listed separately, each marked tested, declared or not
-  supported. It also records which registry models have provider ids, with a date.
-- Universal router docs. The run-4 held-out headline is labeled repository-reported. The
-  shipped prior's refit is documented as reproduced exactly in this repository. The new
-  held-out replay is documented, as are the modeling limits: cascade cost under-predicted by
-  5–22%, optimistic targets, and budgets that bound only expected cost. The dataset pin is
-  corrected to `SWE-bench/SWE-bench_Verified@78f471b`.
-- Research corrections (2026-09-26) in the synthesis, the preprint and the white paper:
+- **`docs/INTEGRATIONS.md` lists what each tool really gets.** For every supported tool,
+  config emission, MCP registration, automatic hooks and enforcement are listed separately,
+  each marked tested, declared or not supported. It also records which registry models have
+  provider ids, with a date.
+- **Universal router docs separate what is measured from what is only reported.** The run-4
+  held-out headline is labeled repository-reported. The shipped prior's refit is documented as
+  reproduced exactly in this repository. The new held-out replay is documented, as are the
+  modeling limits: cascade cost under-predicted by 5–22%, optimistic targets, and budgets that
+  bound only expected cost. The dataset pin is corrected to
+  `SWE-bench/SWE-bench_Verified@78f471b`.
+- **The research papers carry dated corrections (2026-09-26).** In the synthesis, the preprint
+  and the white paper:
   - Theorem D: joint attainability, with a counterexample.
   - The equality condition of the silent-miss bound.
   - "A caught miss is not a completed task."
@@ -169,13 +210,13 @@ evidence behind it. Scripts that read the JSON output may need to adapt:
 
   `research/recompute_corrections.py --theorem-checks` asserts the Theorem D checks without
   data, and the recomputation also prints macro F1.
-- Evidence grades are split into bibliographic verification, claim support, study design,
+- **Evidence grades are split** into bibliographic verification, claim support, study design,
   independent replication and transfer scope. METR's slowdown result is scoped to its
   16-developer, early-2025 study, with a link to the February 2026 update.
-- The Qur'anic lens labels the Arabic source text, the translation, tafsir and the author's
-  design analogy separately, and states what the lens does and does not establish. No Arabic
-  text or translation was changed.
-- The research PDFs are marked as historical, pre-correction editions and recorded in
+- **The Qur'anic lens labels its layers separately:** the Arabic source text, the
+  translation, tafsir and the author's design analogy. It states what the lens does and does
+  not establish. No Arabic text or translation was changed.
+- **The research PDFs are marked as historical, pre-correction editions** and recorded in
   `research/HISTORICAL_EDITIONS.md` (git blob, sha256, pinned commit, figure map, render
   recipe). They were not re-rendered: the Qur'anic text in a fresh render could not be
   verified, and the refutation paper needs a TeX toolchain.
@@ -475,8 +516,11 @@ evidence behind it. Scripts that read the JSON output may need to adapt:
 
 ### Added
 
-- **The universal router's benchmark is independently replicated.** harness-bench run 4 was
-  re-run from the pinned public data (SWE-bench Verified `78f471b`, SWE-bench/experiments
+- **The universal router's benchmark was re-run by the project on a second machine.**
+  _(Corrected 2026-09-26: this entry first said "independently replicated". A re-run by the
+  project itself is not an independent replication, and the held-out harness lives outside this
+  repository; see [docs/UNIVERSAL_ROUTING.md](docs/UNIVERSAL_ROUTING.md).)_ harness-bench run 4
+  was re-run from the pinned public data (SWE-bench Verified `78f471b`, SWE-bench/experiments
   `40f164d`) on a second machine. Results:
   - **Split:** the same 150/350 split.
   - **Held-out test:** 217 of 218 metrics identical, with only wall-clock fit time differing.
@@ -636,10 +680,8 @@ evidence behind it. Scripts that read the JSON output may need to adapt:
 
 ### Fixed
 
-- **A claim minted before the CRLF fold is migrated, not deleted.** Folding `
-` into
-  `
-` changes a claim's content address, so a claim written by an earlier version on a
+- **A claim minted before the CRLF fold is migrated, not deleted.** Folding `\r\n` into
+  `\n` changes a claim's content address, so a claim written by an earlier version on a
   Windows checkout carried the pre-fold address in its filename and failed its own address
   check on load — `loadClaims` returned nothing for it, and `forge ledger verify` reported
   it as an id mismatch. The read path now accepts the pre-fold address as well, so the

@@ -3,7 +3,7 @@
 ## Stack
 
 - Node.js >=20, pure ESM (`"type": "module"`), zero runtime dependencies.
-- Linter/formatter: Biome 2.5.5 (dev dependency).
+- Linter/formatter: Biome 2.5.13 (dev dependency; `npx biome migrate --write` after an upgrade).
 - Types: TypeScript via JSDoc annotations — no `.ts` files, checked by `tsc`.
 
 ## Commands
@@ -24,4 +24,8 @@
 - Run `npm test && npx biome check && npm run typecheck && node src/cli.js docs check`
   before committing — the docs check fails CI when commands/env vars/MCP tools/CHANGELOG
   drift from the code, so update docs IN THE SAME CHANGE, not later.
+- After editing `CHANGELOG.md` (or commands/MCP tools), run `node src/cli.js docs render`:
+  it regenerates the machine-owned blocks, including the Mintlify changelog page
+  (`mintlify/changelog/overview.mdx`), which the docs check fails on when stale. Never edit
+  between `forge:render` markers by hand.
 - Version lives in `package.json` — `scripts/bump.mjs` keeps all manifests in sync.
@@ -1,5 +1,5 @@
 {
-  "$schema": "https://biomejs.dev/schemas/2.5.2/schema.json",
+  "$schema": "https://biomejs.dev/schemas/2.5.13/schema.json",
   "vcs": { "enabled": true, "clientKind": "git", "useIgnoreFile": true },
   "files": {
     "ignoreUnknown": true,
@@ -22,7 +22,7 @@
   "linter": {
     "enabled": true,
     "rules": {
-      "recommended": true,
+      "preset": "recommended",
       "suspicious": {
         "noAssignInExpressions": "off"
       }

@@ -581,7 +581,12 @@ generated from the same registries the check reads, into marker-managed blocks
   every diagram in every tracked markdown file re-themes (deliberate bad examples
   opted out with `docs-check-ignore` are left alone);
 - the repo map in `ARCHITECTURE.md` — drawn from the live import graph, so it cannot
-  drift from the tree it describes.
+  drift from the tree it describes;
+- the docs site's changelog page (`mintlify/changelog/overview.mdx`) — from `CHANGELOG.md`:
+  every release, and the `[Unreleased]` work, as one Mintlify `<Update>` entry listing each
+  change's headline with a link to its full notes. An MDX page cannot hold HTML comments,
+  so its markers are JSX comments (`{/* forge:render:changelog:begin … */}`). A release
+  (`scripts/bump.mjs`) regenerates it in the release commit.
 
 ```console
 $ forge docs render
@@ -594,7 +599,8 @@ $ forge docs render --check
 reconciler in CI: a stale registry-derived block is an **error** whose message is the
 fix (`run forge docs render`), while tree-derived output (the repo map, diagram theme)
 is a warning — moving a file never fails an unrelated PR, but a new command with a
-stale table always does.
+stale table always does. The changelog page is an error too: edit `CHANGELOG.md`, then
+run `forge docs render` in the same change.
 
 ### `forge docs sync` — which prose did this diff make stale?