docs: bind W7 skills seed intake to frozen lifecycle inputs - #1031

Scope

Record the W7 skills seed intake after #1023's released fix, preserving every lifecycle slot, retained seed and original provenance before delta-only facts/fit/role adjudication. This is a source-intake record: no skill selection, installation, comparison amendment, runtime wiring or quality result is claimed.

• PR merge base: d8f7e2051b61bde6361a85b64f41413798015353. Frozen public-input authority: 421bf3f98b631d931838c22a927b4fec246f7df8.
• Lane: lane:foundation.
• Owned paths: docs/decisions/w7-skills-seeds-20261012/; native registrations in manifests/evidence.json.

SOTA sources

• vercel-labs/skills at 7407f389: src/find.ts:87-92 is the vendor's search path; README:87-90 and :132-133 distinguish project scope, named skills and global installation. The central adoption manifest pins CLI 1.7.0. Its historical find execution is unverified because the original output is unretained; no skill add/use/update/check command was executed.
• openai/skills at 49f948fa: README:2 declares deprecation and names openai/plugins. The four installed skills remain frozen copies, each with a separate successor intake; this upstream pointer does not establish a replacement.
• GSD is the existing catalogs/convergence-practice/sources/master-seeds-20261010.json#/instructions-skills/151 seed: open-gsd/gsd-core at ed0f3fd5. Its exact primary metadata remains UNMIRRORED. The archived redirect README at bdcaab2c and the later README at b747aca5 are observations only. Both observed README byte bindings were independently recomputed from public mirror blobs. No new GSD seed, installation, eligible-skill qualification or S1 result is introduced.
• The remaining 56 named source subjects have exact public repository/commit/SKILL.md path and byte-digest bindings in intake.json#/primary_metadata. Metadata name, description and invocation flags are source observations; they are not execution or quality evidence. Another 56 subjects are explicitly UNMIRRORED.
• Frozen landscape inputs at input authority 421bf3f: catalogs/landscape/skills-lifecycle.json, catalogs/convergence-practice/sources/master-seeds-20261010.json, catalogs/landscape/grand-catalog-20261008.json, adoption/skills/manifest.json and catalogs/convergence-practice/public-starred.json. Every seed and slot points back to these published immutable inputs.

Evidence-class table

Local commands run

DISABLE_TELEMETRY=1 npx --yes [email protected] find get-shit-done
reported exit 0; UNVERIFIED historical run (original stdout/stderr unretained); no new probe

Frozen public-input, seed-pointer, slot-count and primary-blob integrity check
exit 0; 5 public inputs, 13 slots, 240 unique master refs,
272 unique grand refs, 153 instructions-skills entries,
56 captured metadata subjects, 56 unmirrored, 0 replacements adopted

python3 scripts/validate.py
exit 0 at 0140a71b; 70 components, 11799 files, 4 profiles, 239 receipts

python3 scripts/evidence_manifest.py --check
exit 0 at 0140a71b; 11799 files

git diff --check
exit 0

gitleaks dir docs/decisions/w7-skills-seeds-20261012 --redact=100 --no-banner
exit 0; no leaks

No runtime/E2E/model test was run. The independent GSD binding check resolves the exact frozen master entry and checks its repository/pin, the primary metadata gap, observation separation and matching README provenance. It fails 7 of 7 assertions at 68b1a00 (exit 1) and passes all 7 at 29f28d4 (exit 0). The broader source-byte, pointer and complete input accounting remain unchanged.

Decision record

docs/decisions/w7-skills-seeds-20261012/README.md, with machine-readable intake beside it. Priority is design-intake, security, ci-pr, skill-lifecycle, then every other lifecycle slot. The lifecycle bytes contain 27 source definitions, 112 slot/source memberships and 30 catalog names; these categories explicitly qualify the addendum's differing 23/115 counts.

The W6 comparison and owner-frozen paired tasks remain the selection authority. The seed lane supplies provenance and gaps; cc-native-practice owns Tier 1 selection, Tier 2 indexing, installation and paired measurement, and the runtime-profile lane owns wiring and non-zero invoke acceptance. Row-36 comparison amendments/execution retain the per-layer W6-LANDED readiness event. This packet does not clear any PENDING source, missing vote, runtime gate or unpublished experiment.

Host evidence

No host receipt or platform-status change.

Follow-ups

• Astra FIRST P3-1 (cited captures; historical discovery explicitly unverified in round 2): attach the exact pinned upstream blob/capture bytes enumerated in REVIEW-EVIDENCE.md and the original stdout/stderr for the [email protected] find get-shit-done discovery, with a digest of published safe bytes, or leave the historical command result explicitly unverified. Mirror-verified metadata does not independently verify that historical discovery execution.

• Astra FIRST P3-2 (historical manifest locators): bind the original commit when reusing inherited numeric manifest-line claims, including the cited Dagu/mise pin-move record and old cross-family review. Preserve frozen historical reviews as history; do not silently turn them into current-head claims. These examples were already stale at the reconstructed base.

• Astra FIRST P3-3 (hostname hygiene): the review counted 18 inherited hostname-containing manifest lines, with none introduced by the diff, new decision files or body. No landing repair is required under the hostname ruling; if those historical records are republished for another purpose, use a public host label under that publication's policy.

• Recover the 56 unmirrored exact skill subjects and their primary metadata at the retained pins.

• Complete pinned metadata coverage for the retained starred registries through the owning Tier 2 index.

• Qualify role/maintenance/effect boundaries and prior verdicts before any S1 delta; retain unknown pins and missing votes explicitly.

• Qualify each frozen-copy successor and the official GSD entry before replacing an installed skill or amending a comparison.

• Run owner-frozen paired tasks and profile invoke acceptance when their named inputs exist; preserve failures and zero invokes.

Checklist

• No new workflow or action dependency.
• No credential, configuration, billing or live profile change.
• Public source-byte digests have published preimages; no private capture authentication is published.
• Peer-owned worktrees, stashes and evidence are preserved.
• Tier/default ownership and readiness boundaries remain explicit.

Claude FIRST P3 Follow-ups

• P3-1 RESOLVED in round 2: the W6 comparison id/path/hash, scope and oracle references are now bound in the intake. The current instructions-skills authority is w6-instructions-skills-20261011, hash f85304bb7bdfac05d361c746babe6a3a62a5186ea06ee72302f4f3039b724774; this intake changes no comparison bytes.
• P3-2: distinguish closeout-triggered S1 and pre-run amendment from execution's sealed source/case/oracle/owner packet. The closeout trigger has fired. Preserve the alternate new-comparison route if the prior comparison has already run; no execution is authorized by this intake.
• P3-3: recover security-audit and loki metadata starting from their bound adoption-manifest paths and SKILL.md digests. Their remaining gap is name/description/flags recovery, not absence of all source bindings.
• P3-4: recover the modelcontextprotocol search-mcp-github subject from the now available 75db1e98 mirror object and preserve its user_invocable spelling. The reviewer fetched the missing promisor pack after intake; later availability does not falsify the original UNMIRRORED observation.
• P3-5 RESOLVED by explicit UNVERIFIED_HISTORICAL_RESULT in round 2; no original discovery output is claimed retained. The reviewer's later native command reproduces the discovery leads and does not supply the original transcript.

The three Astra FIRST P3s above remain carried: cited capture/discovery evidence, historical manifest locators and inherited hostname hygiene. CC correction #187 is reflected in the GSD binding: row 36's add-GSD clause is already met, so the delta-only S1 pass adds no GSD seed.

Round 2 provenance correction

Head 0140a71 fixes all five reproduced review subjects in ONE correction commit, with registry-only rebase of the prior 29f28d4 read head onto main d8f7e20. The non-registry patch IDs of all three authored commits are unchanged by the rebase. W7 S1 remains parked.

• The retrospective resolution snapshot binds all 13 dedicated-layer verdict/wave/sweep gaps and the shared instructions-skills verdict, latest covering sweep, retained returns, cumulative survivor/history, reopen evidence and bounded next action. Absence remains unresolved, never an empty survivor set; the original recovery is not retroactively claimed protocol-compliant.
• Every seed now exposes primary and secondary grand_refs: all 272 unique catalog rows have explicit seed-level provenance. The old per-slot grand_row_refs already listed all 272; the reproduced defect was the seed-level join exposing only 234.
• The GSD regression is in tests/test_w7_skills_seed_bindings.py, alongside guards for catalog joins, resolution state, comparison bindings and probe disposition. The GSD method fails at 68b1a00 (1 test / 1 failure / exit 1); the full five-method suite fails at 29f28d4 (21 failures, no errors / exit 1) and passes after the fix (5 tests / exit 0).
• The unretained historical registry execution is marked UNVERIFIED_HISTORICAL_RESULT, with null verified exit/output references and separately labelled reported results. No replacement probe was run.
• The comparison binds w6-instructions-skills-20261011 at public ancestor 5421e5d (identical bytes at reviewed head 29f28d4), its immutable path/hash, frozen scope, readiness and task/oracle specification references. The actual owner execution packet remains unsealed; no comparison readiness/result is inferred.

The repository guards use CPython unittest and published Git objects through the existing test-discovery path. Source: python/cpython v3.13.16 Lib/unittest/case.py and loader.py, also cited by docs/decisions/2026-10-11-layer-resolution-protocol.md. The parked research passes are not resumed, and no skills or runtime defaults change.