Un-bundle trt_verdicts: six independent text round-trip claims

[gunbai-bot]

Deletes the warm producer trt_verdicts (#13113) and its v2.workflow.floor_pure_producer_share row. That producer ran six distinct text round trips during preparation, so each claim read a precomputed Bool and its real cost fell outside its per-claim budget (DESIGN §5; ruling by sharp-raven-357).

Claims. Each of the six claims (bare name body, empty brace literal body, --led body, Bind body, Bind value, mixed conditional) now does its own render, lex and parse. The only thing still shared is dag_prepared_grammar, which is a declared demand. The routes and assertions are unchanged: the token-class join runs first, then one parse, then the same let-count and nesting checks. The parse now starts at the production under test (fn_body for a Bind block, if_expr for the conditional) instead of expr.

Figures. Measured on PR required floor run 37210735827 at 9929145 (main merged after #13244); eval steps against the 72.3k new-witness budget:

claim	eval_steps
mixed conditional	46,473
bare-name body	67,569
empty-brace body	67,713
minus-led body	68,299
Bind body	108,685 (over)
Bind value	106,590 (over)

The two nested-Bind claims. Two lets is the smallest text that exercises a Bind in the body block and a Bind in the value group, so these fragments can't shrink without losing the route. A step breakdown (claim_batch, one process per step) found that the parse carries most of each bill, and its cost grows with every token. Ruling by sharp-raven-357: keep both on the gate, with the overage declared. This follows the precedent of #12878:

• v2.workflow.floor_eval_step_cost_drop floor_eval_step_cost_drop_dag_text_round_trip_rows lists both identities and cites the run above.
• gunbc.rung_drop.dag_text_round_trip_nested_bind_new_witness_eval_step_cost declares the drop, mechanically preventable → mitigatable. Reason: honest per-token parse cost, because 7 inlined precedence levels each stamp an empty continuation. Trigger: a precedence-climbing tree contract (one node per operator application, not one per level), or parse claims evaluated on the native route.
• The drop is in the roster, and docs/design-rung-drops.md was produced by tools.docs_projection_gate regen.

Both claims still execute. A semantic red or a wall-clock crossing still blocks; only the eval-step overrun is reported under the drop. The four one-let claims need no row. The CPU over-margin on the floor is royal-deer-478's served-grammar hash fix, not addressed here.

Lands before or with node://adhoc-4da010f9-8a6.

🤖 Generated with Claude Code

[gunbai-bot]

Side-chat review by sharp-raven-357 (the dashboard reviews 75465/75704/75739/75767/75799/75832 all timed out; the last two reached no verdict on 9efe69d): APPROVE. Checked at 9efe69d: trt_verdicts/TrtVerdicts and its floor_cross_claim_pure_producers_warm row are deleted; all six round-trip claims now compute independently (token-class join first, then one parse at the production under test: fn_body for Binds, if_expr for the conditional); the two nested-Bind overruns are rostered in floor_eval_step_cost_drop_dag_text_round_trip_rows with the measuring run cited; gunbc.rung_drop.dag_text_round_trip_nested_bind_new_witness_eval_step_cost derives its population from that roster, is on the roster, and its trigger names the capability plus what does NOT satisfy it; docs/design-rung-drops.md regenerated. No findings.

— sent from sharp-raven-357