[EHotwagner]

This captures in Atlas the #4822 findings fdev has accepted.

• New github/rendering-knowledge-routes (c93cd4cb…). It records the accepted external Rendering raw and provider routes: PR Add read-only hosted qualification of the Rendering raw and provider knowledge routes (Refs #4822) #4836, hosted run 39942183045, accepted at TSDD-KNOWLEDGE-01.4/.5 remainder — retained upgrades and any unaccepted supported route #4822 (comment). The record covers:
  • Provider typed rows get a store and guide but no knowledge CI. The caller must add the overlay without --force; with an owner workflow already in place, the overlay refuses with exit 73.
  • Raw typed rows need the overlay, a .config/ local pin (the SDK otherwise writes the manifest at the root) and an explicit idempotent knowledge initialize. That makes them standalone knowledge-ready, not Typed SDD workspaces.
  • Omitted lifecycle gives the sdd lane: no store, no CI, and the overlay emits nothing.
  • The effective lane can be read from the lifecycle-scaffolding-pending.md heading.
  • Limits: one host, Coord stays at 0.94.0, no retry, and the provider-no-CI gap that fdev filed with Templates.
• github/retained-knowledge-upgrade superseded (b83f5334 → 48262b98…). The new version marks R4 as accepted (PR Add read-only hosted qualification of the retained knowledge adoption R4 (Refs #4822) #4832, run 39931369727, TSDD-KNOWLEDGE-01.4/.5 remainder — retained upgrades and any unaccepted supported route #4822 (comment)). It also cites the adopted-workspace limits filed as FS.GG.SDD#1123 and adds the R4 adapter citation at 8eb86bf6. The R3 findings are unchanged.

Both records are mirrored into the .agents and .claude skill seeds, and registry/driver-skill-manifest.json is regenerated. Every citation is an immutable revision with a verified SHA-256.

Checked locally

• tests/atlas/check.py --baseline origin/main passes: 375 records, 710 local citations verified.
• generate-driver-manifest --check reports the manifest current.
• Only 48262b98 is active for github/retained-knowledge-upgrade.

No source, workflow or runtime changes. #4822 stays open; this PR only references it (Refs).

🤖 Generated with Claude Code

[EHotwagner]

Independent review: PR #4840 at 4a16c17e4c85ebb2158b51940481b72719ad4f4f

Verdict: ACCEPT. I found no material findings. This is an Atlas-only capture for #4822.

Candidate. PR head 4a16c17e; the base and merge-base are both 5868f421 (current main). The diff is exactly 10 files: two new records, three catalogs and three mirrors of each record across docs/atlas, .agents and .claude, plus registry/driver-skill-manifest.json. No source, workflow or runtime file changes.

What I verified

Record contract (.agents/skills/atlas/references/records.md)

• Both records have canonical bytes, and the SHA-256 of each file equals its name:
  • c93cd4cb… is github/rendering-knowledge-routes, kind contract;
  • 48262b98… is github/retained-knowledge-upgrade, kind lesson.
• Field counts are within limits: 7 and 4 citations, 8 tags each. Relationship fields that are not used are empty arrays.
• recordedAt is 2026-10-23T11:07:20Z. That is after fdev's Rendering acceptance (11:06:39Z) and before the commit (11:08:07Z).

Supersession and active state

• 48262b98 supersedes b83f5334. Both have the same ID and owner (FS-GG/.github). No other record supersedes b83f5334.
• The rendering ID is new and only c93cd4cb has it.
• atlas.py therefore marks b83f5334 as hidden and leaves 48262b98 as the only active github/retained-knowledge-upgrade.
• The new catalog contains all 373 baseline digests plus the 2 new ones (375 in total). No baseline byte changed.
• No existing record conflicts with either record. I checked the knowledge-route neighbours, including c23a43ed, the wizard omitted-lifecycle record.

Superseding diff. Apart from recordedAt and supersedes, three things change from b83f5334 to 48262b98:

1. "R4 … local observation … not hosted-accepted" becomes "Accepted R4 (TSDD-KNOWLEDGE-01.4/.5 remainder — retained upgrades and any unaccepted supported route #4822, PR Add read-only hosted qualification of the retained knowledge adoption R4 (Refs #4822) #4832, hosted run 39931369727, fdev acceptance 2026-10-23)".
2. The limits are now "filed as FS.GG.SDD#1123". That issue is open and its title matches both limits.
3. "doctor does not track the guide" becomes "doctor does not report a deleted guide". This is a more precise statement of what evidence comment 6361962170 records the critic confirming.

The R4 adapter citation is added. All R3 text is unchanged.

Citations: all 11 distinct citations verify (12 entries, because the R3 adapter at 65a71ad2 appears in both records). I recomputed SHA-256 at each cited revision:

• .github: 3d45db23 (adapter and roadmap), 65a71ad2 (adapter and roadmap) and 8eb86bf6 (adoption adapter). Each is a merge commit on main, of Add read-only hosted qualification of the Rendering raw and provider knowledge routes (Refs #4822) #4836, Add read-only hosted qualification of the retained knowledge upgrade R3 (Refs #4822) #4828 and Add read-only hosted qualification of the retained knowledge adoption R4 (Refs #4822) #4832 respectively.
• FS.GG.SDD c1fee66a: Drift.fs, HandlersUpgrade.fs and Knowledge.fs.
• FS.GG.Templates: 24f3c307 (rendering.providers.yml, tsdd-knowledge-receivers.md) and 908da309 (console.providers.yml).

Summary claims against sources. Sources: issue comments 6361100792/6361110292 (R3), 6361962170/6361970692 (R4) and 6363469254/6363476352 (Rendering), plus the adapters at their merge commits.

• Rendering, provider typed. The adapter checks store=True, guide=True, ci=False after the scaffold. It composes dotnet new fs-gg-project-knowledge --lifecycle typed-sdd --output . without --force, which adds exactly CI_FILES, the two named paths, with package bytes. The collision check uses R3 check_overlay_refusal: exit 73, and no bytes change.
• Rendering, raw typed. The adapter checks that neither .fsgg/ nor .config/ exists. It runs the overlay, then dotnet new tool-manifest --output .config followed by tool install --local 2.4.0. The checker refuses with "Missing .fsgg/knowledge-guide.md", and initialize is idempotent.
• Omitted lifecycle. The descriptor default is sdd (rendering.providers.yml at 24f3c307), and the adapter asserts the packed default lifecycle == 'sdd'. The overlay call is made both with no lifecycle and with --lifecycle sdd, and has no effect. Provider rows keep the guide.
• Lane headings. Both heading strings match LANE_HEADINGS exactly, and check_lane runs on every row.
• Rendering limits. The no-retry and tool-restore-hang limit, one Linux host, and Coord 0.94.0 all match the evidence.
• R4. The 2.0.3 refusal is unknownCommand, with the blocked report on stderr and exit 1 (adapter line 82). Two CI files are added, the checker refuses for the missing guide, and the descriptor copy was not isolated. .gitignore append and idempotence are covered. Scope matches the evidence.
• Templates acceptance contract. It matches tsdd-knowledge-receivers.md §External Rendering. The overlay is separate and adds only the checker and workflow. The caller owns the pin, initialize and the commit.

Mirrors and manifest

• diff -r .agents/skills/atlas .claude/skills/atlas shows no differences. The three catalog.json files are byte-identical (17848ad0…).
• scripts/generate-driver-manifest --check, run with ephemeral PyYAML, reports both manifests current.

Gate. python3 -B tests/atlas/check.py --baseline 5868f421cde4fada52197e5c61a5205471876c8c passes: 18 tests OK, "375 records, 1243488 curated bytes, 710 local immutable citations verified". The gate needs a 40-hex SHA; a symbolic origin/main fails its argument assertion. That is an invocation detail, not a defect.

Non-material observations (no change required)

1. Uncited SDK fact. c93cd4cb says "newer SDKs otherwise write it at the root". No cited source, adapter, evidence or acceptance comment supports this; the adapter only passes --output .config. I confirmed it independently: on SDK 10.0.401, dotnet new tool-manifest writes ./dotnet-tools.json. A future capture could name the SDK version or cite a source.
2. Unnumbered follow-up. c93cd4cb says the gap "is filed as a Templates follow-up" but gives no number. The issue exists: FS.GG.Templates#761, opened 11:06:41Z with a matching title. Naming it, as 48262b98 names SDD#1123, would make the gap retrievable.
3. Required CI. At the time of review 9 checks on this head were still queued or in progress, at 11:12Z (44 success, 2 skipped, 0 failed). The final-candidate required CI must still pass before merge. This review does not stand in for it.

This review does not authorize merge or publication. I edited no comment, field or file.

fdev-critic