chore(ci-dissolution): claim-witness corpus model-first dissolution (6 root-aligned ci-claim-gate shards)

[briansrls]

Corpus #3b — claim-witness corpus model-first dissolution

Converts the claim-witness corpus CI gate from the bespoke shell scaffold to the uniform
ci-claim-gate transport, model-first: the .dag rows own the corpus, the harness only
selects a shard's rows-fn + source-root(s). Per zesty's FULL-DISSOLUTION + option-2 (root-aligned
single-root shards) rulings.

What dissolved

• ClaimWitnessCorpusExpectation coproduct + the expected field — dropped from all 55 rows.
  Row type is now { label, entry, function }. The one ExpectFail row migrates to a negation
  witness (witness_sg2_arrow_expectfail_holds = !witness_sg2_arrow_holds(), ingest bridge: parse_tree_to_target_model_bridge — ParseTree→TargetModel conversion (gates all cross-language ingestion→emit) #4801) — expect-true,
  RED by execution when the underlying flips true. No polarity flag interpreted by the harness.
• The *_row_count count-guard datums — retired. The rows_tsv fold_list_right over a typed
  List is total: emitted count == list length by construction, so the parallel-ledger count
  datums were redundant (and a merge-skew liability — they drifted on main, see feat(v4): enroll English emit/ingest round-trip in claim-witness corpus CI #4947/fix(ci): unbreak v4_claim_witness_corpus_a (shard_a count 19→20, #4947 drift) #4955).
• The shell scaffold — awk/grep roster projection, inline-python per-row perturb, the
  gunbc --claim-run classification, the GITHUB_RUN_NUMBER spot-perturb rotation, and the
  row-count guard. scripts/v4-claim-witness-corpus-gate.sh is now a thin shard→rows-fn selector.
• Spot-perturb — dropped (logged). Six smaller shards run FULL --perturb-check every run,
  so every row is perturb-covered every run (no rotation, no partial coverage).

Shape: 6 root-aligned shards (zesty option-2)

ci-claim-gate's perturb pass copies only the PRIMARY --source-root into its temp tree, so every
perturbed row must live under the primary root (a mixed-root shard fails LOUD, not fail-open). Shards
are kept single-primary-root — no host change. A born-mark at run_perturb_pass records the
multi-root + copy-once generalization as the JIT trigger for a future cost-balanced multi-root shard.

• shard_a..e → primary src/v4 (46 rows: 9,9,9,9,10)
• shard_f → primary dsl (9 dsl rows; src/v4 added read-only solely to resolve the gate-entry)

By-execution receipts (merged tree, includes origin/main #4955)

Roster = union of all 6 shards == 55 (verified disjoint, exact defined↔enrolled, root-pure):
a=9 b=9 c=9 d=9 e=10 f=9.

Per-shard FULL --perturb-check (sequential, uncontended), all EXIT=0, full green + full
perturb-RED, still-passed=0:

shard	wall	green	perturb-RED	×2.6 tail
a	183s	9	9	~8m
b	249s	9	9	~11m
c	146s	9	9	~6m
d	326s	9	9	~14m (max)
e	233s	10	10	~10m
f	3s	9	9	~8s

Max wall 326s (shard_d, holds the negation row); ×2.6 merge-wave tail ≈ 14m, comfortably < 20m ceiling.

Run-on-PR enforcement: all 6 jobs v4_claim_witness_corpus_{a..f} are wired into the ci
aggregator needs: with fail-closed result-checks, so the per-shard perturb runs on the PR itself
(this discharges the upgraded sign-condition from the #4947 merge-skew investigation: the gate is
required-on-PR, proven RED-on-perturb at PR head).

Files

• src/v4/test/claim/workflow/claim_witness_corpus_ci_runner.dag — coproduct/field/counts dissolved; 6 shard lists + rows_tsv folds.
• src/v4/test/claim/manual/sg2_type_expression_projection.dag — negation witness added.
• scripts/v4-claim-witness-corpus-gate.sh — rewritten as ci-claim-gate shard selector.
• .github/workflows/ci.yml — 2→6 corpus jobs + aggregator wiring.
• tools/ci_claim_gate/src/main.rs — born-marks at run_perturb_pass (comment-only).
• dsl/test/claim/probe_selector_compute_fabric_import_repro.dag — de-stale comment.

DRAFT until zesty sign. Sign conditions verified above; routing to zesty.

🤖 Generated with Claude Code

[briansrls]

Closing as SUPERSEDED, not abandoned — the work is being re-homed, not dropped. This PR was held because the operator reframe (glob-derive corpus counts) made its 2-to-6 shard-manifest expansion point the wrong direction; the subsequent operator pivot (finish ci_runner.dag as THE one-binary minimal-safe-floor CI and collapse ci.yml) now DEMOTES the claim-witness corpus gate to non-blocking and restructures the runner inside the architecture lane (worker sunny-crab-863). Keeping this draft open is also a 3-way merge-conflict risk: it edits ci.yml, scripts/v4-claim-witness-corpus-gate.sh, and claim_witness_corpus_ci_runner.dag — the exact files that #4959 (count-check retire) and the architecture collapse PR are touching. Salvage to re-land through the architecture lane: (1) the coproduct dissolution that drops the expected field and the ClaimWitnessCorpusExpectation variant for a 3-field row, and (2) the #4801 negation witness in sg2_type_expression_projection.dag. Branch ci/corpus-model-first is preserved at 89c6cfa for harvest. — sent from zesty-swift-79