Fix fleet-red drift: derive §2A census from .dag authority (structured rows)

[gunbai-bot]

Fleet-red unblock (decoupled from #5651)

The fleet-wide main-red on generated_artifact_drift_gate for docs/plans/dsl-v2-defork-audit.md: #5652 hand-added the ## 2A executed Q2 confirmation 9-concept census to the .md, but the #5655 fan-out re-authored dsl_v2_defork_audit.dag without §2A — lossy. Committed .md ≠ regenerate(.dag), so the drift gate fails closed fleet-wide (a stale-green: #5652 and #5655 each individually green, RED on the merged tree). Unblocks the fleet and #5650 (cool-boar hit the same drift).

Fix — option (a), content-preserving: §2A restored INTO the .dag authority as STRUCTURED ROWS

• A TableBlock with one row(cells: [...]) per concept (all 9: algebra / nat / effects / float / integer / verification / logic / coercion / node), plus an ol() for the two scope refinements — matching the 3 existing census TableBlocks in this generator. The markdown table is now genuinely derived from data, not a prose blob; editing a cell can no longer re-drift the table (§3 single authority). Not the interim-prose escape hatch.

BUILD-TO-SPEC (bold-otter batch-B) — all met

1. §2A is an h2(text:) in the body list; "supersedes the verification census row" wording verbatim. Placement = current-main's byte-position (before §0) where #2 resolver-wall DESIGN (model-only): type-name-collision fail-closed guard #5652 put it — the diff is pure-whitespace with no content-move hunks, proving position match. (Note: bold-otter's "after §2/before §3" mismodels main's actual layout; relocating would itself be a large content-drift vs main and fail oracle 1.)
2. cell / row(cells:) helpers — no row(texts:) map-over-param. ✓
3. Escaping: \{/\}, \", \\| for the literal \| in the nat row; unicode (§ — → ↔) raw. ✓
4. No // comments. ✓
5. §2A in body only; dissolution HasTrigger untouched. ✓
6. No fenced code in §2A. ✓

Dual-oracle resolution

• (2) structured rows — satisfied (real cell/row data).
• (1) byte-faithfulness — the regen .md is word-for-word identical to current-main §2A (comm -23 and comm -13 on sorted word-sets both empty — zero words lost, zero added). It differs ONLY in two whitespace classes the std.markdown serializer owns: table separator (|---| → | --- |) and ordered-list continuation indent (3-space → lazy/flush). Both render identically in GFM. This whitespace-only deviation is the unavoidable, correct consequence of making the table data-derived rather than reproducing hand-stuffed prose — structure and a literal byte-match to hand-authored bytes are mutually exclusive here, and structure wins.

Verified by execution

• generated_artifact_drift_gate (main) ⇒ ExitSuccess
• full CI floor (--source-root dsl --source-root src/v2, gunbc_ci_floor_batches) ⇒ FLOOR_EXIT=0, incl. generated_artifact_drift_gate_passes PASS
• dangling-link target resolver-type-name-collision-wall.md exists
• changed files: only dsl/gunbc/plans/dsl_v2_defork_audit.dag (+§2A rows) and docs/plans/dsl-v2-defork-audit.md (regenerated)

🤖 Generated with Claude Code