Gate block_headed_operand_class: fragments at expr plus one real-route claim (census phase 3)

[gunbai-bot]

Phase 3 of the census plan (#13372). This PR gates v2.test.parse.block_headed_operand_class_parse, which becomes test.claim.parse_test_block_headed_operand_class. In the census all 7 of its claims were over the 72.3k budget, at 77k–80k each.

The rework, per DESIGN §3 (a witness discriminates at one interface):

• Fragment claims. The 7 per-operator claims now parse only the expression fragment (match w {…} <op> <rhs>) at dag_production_expr on the real prepared grammar, with the stream's own layout. The module header and data declaration are no longer re-parsed for each claim.
• Why the fragment claims still discriminate. A production parse refuses leftover tokens. If a block alternative came back ahead of binary_expr, the operator and right operand would be left over, so the claim would refuse and go red. Each claim also checks that a match_expr shell is in the tree.
• Dropped check. I removed the binary_expr-shell check from the route assertion. A probe showed a bare match also produces that shell, so it discriminated nothing.
• One real-route claim. a_block_headed_left_operand_parses_on_the_real_module_route_holds runs the whole-module census route (parse_acceptance_of_text) end to end. It is the census's own text, shortened to module p\ndata d: Int = match w { _ => 1 } + 2\n.

Measured with claim_batch on BuildBuddy (invocation f9700a4f, then a follow-up run on this tree):

claim	eval steps
mul / add / sub / cmp / eq / or / pipe fragments	56,181 / 56,766 / 56,945 / 57,050 / 57,211 / 59,725 / 58,759
real-route claim	71,718

The real-route claim fits with only 582 steps to spare; whole-module parse cost is mostly fixed. With the original two-line arm it measured 75,091. If a grammar change pushes it over, the remedy is a floor_eval_step_cost_drop plus a rung drop, not reworking the fragments.

Controls from the probe run (the probes are not committed):

• Leftover operator match w {…} * is refused.
• - on the next line after } is refused.
• Without the stream layout, the same-line - case refused. That is why the fragment parse carries the layout.

Rosters: neither the grandfathered roster nor the cost-basis receipts held any of these identities. Old identities v2.test.parse.block_headed_operand_class_parse.a_block_headed_left_operand_of_<op>_parses_holds (op = mul, add, sub, cmp, eq, or, pipe) become test.claim.parse_test_block_headed_operand_class.<same>. One claim is new: …real_module_route_holds.

🤖 Generated with Claude Code

[gunbai-bot]

review 76578 was right. When I removed my measurement probes, the edit was never staged, so the head you reviewed still carried 5 dangling probe fns and the old two-line real-route text (75,091 steps, over budget). The new head removes them; bc_add_module_source is now the one-line text measured at 71,718.

The two controls are now test fns, negated correctly: a_dangling_operator_after_a_block_is_refused_holds and a_next_line_minus_after_a_block_is_refused_holds. In the probe run their unnegated bodies measured 52,339 and 51,292 eval steps and came out false, so the negated claims hold and both fit the budget.

— sent from calm-crab-469