[EHotwagner]

This adds the Atlas capture fdev requested on #4822 (#4822 (comment)).

• New github/wizard-022-knowledge-rows (74b5d2ed…). It records the accepted Wizard 0.22.0 rows from PR Add read-only hosted qualification of the published Wizard 0.22.0 knowledge rows (Refs #4822) #4841 (run 39948229198):
  • When the lifecycle is omitted, fable-game takes the descriptor default typed-sdd (the Wizard logs "selected from published provider"). Rendering and every other template take the Wizard's own sdd.
  • The Rendering default rows give the sdd lane: no store, no knowledge CI.
  • fable-game PLAYER and COMPLETE give identical typed workspaces whether the lifecycle is explicit or omitted. Pins, owner CI and the full knowledge contract all pass.
  • The non-SVG refusal control is ratified (--svg-foundation exits 2 with "unknown argument"). Non-SVG readiness is covered by [TSDD-KNOWLEDGE-01.4/.5] Qualify published Fable-game non-SVG provider automatic typed knowledge #4773 and [TSDD-KNOWLEDGE-01.4/.5] Qualify RAW legacy non-SVG Fable typed knowledge and omitted Standard preservation #4769.
  • The complementary pinned Templates evidence is cited: Templates#760, run 39946080162.
  • Limits of the unpinned run are recorded.
• github/rendering-knowledge-routes superseded (c93cd4cb → daa0fc9f…). The only change names the provider no-CI follow-up as FS.GG.Templates#761 (the earlier critic's note). Everything else is byte-identical.

Both records are mirrored into the .agents and .claude seeds, and registry/driver-skill-manifest.json is regenerated. Citations are pinned to immutable revisions with verified SHA-256: the adapter and roadmap at 1e8299ae, the Wizard Program.fs at its source 6e6fb9b9, and the Templates descriptors at 24f3c307.

Head c770eb32 rebuilds the same change on main 1fac8e92. #4838's Atlas/registry capture made the earlier head conflict. The record bytes and digests are unchanged (74b5d2ed…, daa0fc9f…); only catalog.json and the driver manifest were regenerated on the new base.

Checked locally:

• tests/atlas/check.py --baseline <origin/main sha> passes (379 records, 721 local citations verified).
• generate-driver-manifest --check reports the manifest as current.
• Each ID has exactly one active record.

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

🤖 Generated with Claude Code

[EHotwagner]

Independent review: PR #4844 at 92898bc217d7765c789666c2e7bef05e8724a0d6

Verdict: ACCEPT. This review covers the exact head above. If the head changes, this decision no longer applies.

Scope: the Atlas capture for #4822. It adds github/wizard-022-knowledge-rows (74b5d2ed…), supersedes github/rendering-knowledge-routes (c93cd4cb → daa0fc9f…), mirrors both into the .agents and .claude seeds, and regenerates registry/driver-skill-manifest.json. The base is c7297c11, which is current main. The 10 files in the PR are the 10 files in the local diff.

Material findings

None.

Verified

• Record contract (.agents/skills/atlas/references/records.md): all required fields are present, with kind = contract, status = recorded, 9 tags, 5 citations, and empty relationship arrays where nothing applies. Each file's sha256 equals its name. recordedAt (12:00:47Z) is after fdev's acceptance (12:00:13Z).

• Supersession is limited to the one sentence. Compared field by field with c93cd4cb, only recordedAt, supersedes (= [c93cd4cb…]) and one summary sentence differ. That sentence changed from "the gap is filed as a Templates follow-up" to "the gap is filed as FS.GG.Templates#761". Templates#761 ("Typed Rendering via provider gets knowledge store but no knowledge CI…", open) is that exact gap and was filed from run 39942183045. Citations, tags, title and kind are unchanged. The supersession has the same ID and the same owner.

• One active record per ID: the catalog has 377 records, sorted, with no orphan files. Every ID has exactly one active record. github/rendering-knowledge-routes → daa0fc9f is active and c93cd4cb is superseded. github/wizard-022-knowledge-rows → 74b5d2ed only. No supersedes or conflictsWith reference dangles. No baseline digest was removed.

• Citations: I recomputed every citation in both records from Git objects. Five of them are new:

  • .github: 1e8299ae adapter 810be996… and roadmap 18956211…, and 6e6fb9b9 Program.fs e8d4556b…;
  • FS.GG.Templates 24f3c307: fable-game.providers.yml 6b3d8f2d… and rendering.providers.yml ff49732a….

  The rendering record's inherited 3d45db23 and 24f3c307 citations also match. 24f3c307 is an ancestor of Templates origin/main.

• Mirrors and manifest: docs/atlas, .agents/skills/atlas/assets/seed and .claude/skills/atlas/assets/seed have identical records and catalogs. scripts/generate-driver-manifest --check passes ("driver-skill-manifest.json and coordination-kit-skill-manifest.json are current"; projections current).

• Atlas gate: python3 -B tests/atlas/check.py --baseline c7297c11e5f68e363d1615dcfce4ada00b64257f exits 0 with "377 records, 1250888 curated bytes, 713 local immutable citations verified".

Summary statements checked against sources

Sources: evidence comment 6364363770, fdev's acceptance and ratification 6364370377 (both signed fdev), the adapter at 1e8299ae, Program.fs at 6e6fb9b9, and the descriptors at 24f3c307.

• Identities match the evidence: 0.22.0 archive caa91285 with source 6e6fb9b9, SDD 2.4.0, Workspace.Template 0.19.3, UI 0.32.3, run 39948229198 and PR Add read-only hosted qualification of the published Wizard 0.22.0 knowledge rows (Refs #4822) #4841.
• fable-game omitted lifecycle takes the descriptor default; every other template takes sdd. selectScaffoldLifecycle applies the declared default only when Template = "fable-game" and --lifecycle was not given. Otherwise it uses opts.Lifecycle, which the parser initialises to "sdd". The fable-game descriptor declares default: typed-sdd and the Rendering descriptor declares default: sdd.
• Log labels: header prints "selected from published provider" for an omitted fable-game lifecycle and opts.Lifecycle otherwise. This matches the raw-output rows in the evidence.
• Rendering default rows: sdd lane, no store, no knowledge CI, guide seeded. The adapter checks this with check_knowledge_state(store=False, guide=True, ci=False) and check_lane.
• fable rows:
  • The pins match the adapter's FABLE_TOOLS.
  • The owner CI is compared byte for byte against the 0.19.3 fs-gg-fable-game payload.
  • The initial commit, Qualify remaining public knowledge routes and retained FSharp compatibility migration #4723 probe, ALL-shared export after private capture, 10,485,760/10,485,761 boundary and preservation all appear in both the adapter and the evidence.
  • The default bundle is player.
• No background commit: "Initialization required … The wizard does not run initialization" is at Program.fs line 1854, and the evidence confirms it.
• Non-SVG: the CLI parser has no svg option and no general --param passthrough. The parameters forwarded to fsgg-sdd scaffold are a fixed set. other :: _ -> Error "unknown argument: …" is reached before run, and the parse-error path returns 2. The adapter requires exit 2, that message and no target. fdev ratified the substitution, and [TSDD-KNOWLEDGE-01.4/.5] Qualify published Fable-game non-SVG provider automatic typed knowledge #4773/[TSDD-KNOWLEDGE-01.4/.5] Qualify RAW legacy non-SVG Fable typed knowledge and omitted Standard preservation #4769 are cited for coverage, as in the acceptance.
• Limits: unpinned dependency on the Templates main descriptors (hashes match DESCRIPTOR_SHA), a no-op self-update, and fail-closed behaviour on a later publish. All match evidence item 3 and the acceptance.

Nonmaterial observations (no change required)

1. "fable-game PLAYER … and COMPLETE, with explicit typed-sdd or omitted lifecycle, give identical typed workspaces" can be read as PLAYER being identical to COMPLETE. The evidence says the explicit and omitted rows are identical within each bundle (258 unchanged files for PLAYER, 325 for COMPLETE). The next sentence ("Each has …") points to the right reading.
2. "Pinned Wizard creations with product builds are separately qualified by … Templates#760, run 39946080162" states the Templates result as fact. The evidence says the .github owner cited that run but did not review it, and the acceptance only says "is cited". A tighter wording would be "covered by the Templates-owned verifier (not reviewed here)".
3. The older lesson github/wizard-omitted-lifecycle-descriptor-default (c23a43ed) ends "Source delivery does not publish Wizard 0.22.0 …". That is historical and does not conflict, so leaving it unsuperseded is acceptable.

Not covered by this review

• When I read it, required CI on 92898bc2 was still running (31 success, 1 skipped, 19 queued or in progress). It must finish green before merge.
• I did not re-download the hosted run artifact. The row facts rely on the accepted evidence and the second artifact review.

fdev-critic

[EHotwagner]

Independent review: PR #4844 at c770eb32d4c8f2e6d5035019be52d2aaf4d8f8c0 (rebased onto 1fac8e92)

Verdict: ACCEPT. This review covers the exact head above and replaces my decision for 92898bc2 (record). If the head changes again, this decision no longer applies.

Material findings

None.

Verified at this head

• Base and scope: the parent of c770eb32 is 1fac8e92, which is current main and includes Reconcile accepted Creator 0.22 publication in the registry and Atlas (Refs #4811) #4838. Against main the PR touches the same 10 files as before: two records times three roots, three catalogs, and registry/driver-skill-manifest.json. Nothing else changes.
• Record files are unchanged from 92898bc2: I compared the Git blob IDs. 74b5d2ed… and daa0fc9f… are byte-identical to 92898bc2 in docs/atlas, .agents/skills/atlas/assets/seed and .claude/skills/atlas/assets/seed. My content findings on 92898bc2 therefore carry over unchanged. They covered the summary checked against evidence comment 6364363770 and acceptance 6364370377, the adapter at 1e8299ae, Program.fs at 6e6fb9b9 and the descriptors at 24f3c307, plus the one-sentence Templates#761 supersession and the recomputed citations.
• Catalogs: docs/atlas/catalog.json is sorted with 379 records. That is all 377 main records, none removed, plus exactly 74b5d2ed and daa0fc9f. The record files match the catalog one for one. The .agents and .claude seed catalogs and record directories are identical to docs/atlas.
• One active record per ID: every one of the 379 records' IDs has exactly one active record, and no supersedes or conflicts link points at a missing record.
  • Reconcile accepted Creator 0.22 publication in the registry and Atlas (Refs #4811) #4838 added 666e7f69 (contract/new-sdd-workspace, which supersedes 09e90ad7) and 9d1a0da5 (github/creator022-publication-boundary). Each is its ID's only active record.
  • github/rendering-knowledge-routes resolves to daa0fc9f, with c93cd4cb superseded. github/wizard-022-knowledge-rows resolves to 74b5d2ed.
  • Reconcile accepted Creator 0.22 publication in the registry and Atlas (Refs #4811) #4838's records agree with the new record on Wizard 0.22.0, archive caa91285 and source 6e6fb9b9. Neither contradicts the other.
• Manifest: compared with main, the diff is only the atlas tree-sha256, the seed catalog sha256 (ef889315…, which equals the committed seed catalog) and the two added record entries.
• Gates, run from a scratch detached checkout of c770eb32 that has since been removed:
  • python3 -B tests/atlas/check.py --baseline 1fac8e920392b9dca50b8ae315bcadb36216cd3f exits 0: "379 records, 1256760 curated bytes, 721 local immutable citations verified".
  • scripts/generate-driver-manifest --check exits 0: both manifests are current and the projections are current. It ran with the already-built engine, and src/ is unchanged since c7297c11.

Nonmaterial observations

These carry over from the previous record unchanged: the PLAYER/COMPLETE "identical" wording, the unreviewed Templates#760 claim, and the historical sentence in c23a43ed.

Not covered by this review

• Required CI on c770eb32 had not finished when I read it: 47 queued, 4 in progress, 2 success, and 2 cancelled (fixture, declared-closes). The merge state is BLOCKED. The cancellations may be concurrency supersession, but I did not confirm that.
• Required CI must finish green on this head, and any cancelled required check must be rerun, before merge.

fdev-critic