foundation: retain GPT-only initial model-routes sweep - #1177

Single-family (Claude) initial sweep of foundation/model-routes under the
CC's 2026-10-15T19:52Z ruling (A). Three landscape-sweep-worker stages ran on
Claude Opus 5.5 at effort max, staged by the harness's own build_args.py,
with a hand-built layer input because the layer is unregistered at
8053f14. 19 proposals (14 carried + 5 new), facts refuted 0, Claude fit
refuted 8, 0 survivors in this half (the GPT fit vote is missing by design,
so 11 are refuted only by absence). The critic's flagged follow-up was not
run under the $40 cap and is recorded as a retained failure. No catalog,
convergence or ledger file is edited; the merge with the GPT half (#1177)
follows MERGE-PLAN.md and supports TRIAL labels only.

Corrections from the CC's full read (verdict MICRO-full-afcc08cb.md):
- No digest of unpublished bytes remains (correction #161). The digests of
  the unmasked layer input, the three stage scripts as run, the merged
  proposal list and the CC direction record are removed. hashes.json keeps
  only published or recomputable values and names the private originals
  under unpublished_originals without any digest. The direction record and
  the Muse re-qualification receipt are cited by name only, and no
  abbreviated digest is left.
- The published decision_ref names the direction record without its
  digest, so requirement_sha256 is recomputed over the published fields
  with scripts/saturation_ledger.py requirement_sha256 (deviation 7).
- returns.json is reconverted with convert.py --usage over the three stage
  usage records, which join_usage.py joins unchanged into usage-joined.json
  (deviation 8). All 38 Claude votes now carry the measured claude-opus-5-5
  at effort max, and no effort_deviation or web_search_capped failure is
  retained. The counts are unchanged.
- The cost estimate states the 1-hour cache TTL as its assumption
  (prompt-caching.md:144 gives the rate for each TTL). MERGE-PLAN.md cites
  recipes/saturation-sweep.md:99 and :174-175 and labels the
  vendor-document capture a PROPOSAL.
- tests/test_provenance_digest_custody.py adds
  ModelRoutesClaudeHalfCustodyTests. Every 64-hex value in the record's two
  directories must hash a tracked file (as checked out or at 8053f14) or
  be prompts_sha256, requirement_sha256 or platform_profiles_sha256
  recomputed with the harness's own code at 8053f14. Private records are
  cited by name, and no abbreviated digest stands in for a value. The class
  fails on the first head's tree and passes here.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Single-family (Claude) initial sweep of foundation/model-routes under the
CC's 2026-10-15T19:52Z ruling (A). Three landscape-sweep-worker stages ran on
Claude Opus 5.5 at effort max, staged by the harness's own build_args.py,
with a hand-built layer input because the layer is unregistered at
8053f14. 19 proposals (14 carried + 5 new), facts refuted 0, Claude fit
refuted 8, 0 survivors in this half (the GPT fit vote is missing by design,
so 11 are refuted only by absence). The critic's flagged follow-up was not
run under the $40 cap and is recorded as a retained failure. No catalog,
convergence or ledger file is edited; the merge with the GPT half (#1177)
follows MERGE-PLAN.md and supports TRIAL labels only.

Corrections from the CC's full read (verdict MICRO-full-afcc08cb.md):
- No digest of unpublished bytes remains (correction #161). The digests of
  the unmasked layer input, the three stage scripts as run, the merged
  proposal list and the CC direction record are removed. hashes.json keeps
  only published or recomputable values and names the private originals
  under unpublished_originals without any digest. The direction record and
  the Muse re-qualification receipt are cited by name only, and no
  abbreviated digest is left.
- The published decision_ref names the direction record without its
  digest, so requirement_sha256 is recomputed over the published fields
  with scripts/saturation_ledger.py requirement_sha256 (deviation 7).
- returns.json is reconverted with convert.py --usage over the three stage
  usage records, which join_usage.py joins unchanged into usage-joined.json
  (deviation 8). All 38 Claude votes now carry the measured claude-opus-5-5
  at effort max, and no effort_deviation or web_search_capped failure is
  retained. The counts are unchanged.
- The cost estimate states the 1-hour cache TTL as its assumption
  (prompt-caching.md:144 gives the rate for each TTL). MERGE-PLAN.md cites
  recipes/saturation-sweep.md:99 and :174-175 and labels the
  vendor-document capture a PROPOSAL.
- tests/test_provenance_digest_custody.py adds
  ModelRoutesClaudeHalfCustodyTests. Every 64-hex value in the record's two
  directories must hash a tracked file (as checked out or at 8053f14) or
  be prompts_sha256, requirement_sha256 or platform_profiles_sha256
  recomputed with the harness's own code at 8053f14. Private records are
  cited by name, and no abbreviated digest stands in for a value. The class
  fails on the first head's tree and passes here.

Co-authored-by: Claude Opus 5.5 <[email protected]>