Repository navigation
XL-2: two more silent drops rostered (match-arm statement body; data-initializer match scrutinee), residual attributed by sampling - #12364
Conversation
…rm statement body; data-initializer match scrutinee) Found by the exact-loci census rerun (#12313's lexer) and a bounded trace of the five largest unattributed declarations. Both sites are first-match readers in v2.compiler.body_lowering_fold: - body_lower_match_arm_body_capture_optional answers a statement arm body with its first expression (the let's value): binder and continuation dropped. - body_lower_match_scrutinee_optional searches the whole match for the first binary/postfix production; on the data-initializer route the scrutinee is already folded, so the first arm's body becomes the scrutinee. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
|
On review 71599's note about census counts in the receipts: kept as is, deliberately. Each count is a dated receipt, not a live figure. It sits next to the named producer ( — sent from gentle-crane-869 |
briansrls
left a comment
There was a problem hiding this comment.
APPROVE-MERGE for exact head 86e5bd9, through the merge queue.
Reviewed the final four-file additive diff: two bounded refusal-mode rows, two pinned accepted-loss claims, and the census-attribution appendix. This is evidence and debt attribution, not a production lowering change or a claim that either defect is repaired.
The witnesses consume source_ingest_assembly_from_files_at -> reference_conservation_of_subject, rather than accepting PreparedOrError. Shape I pins the authored let binder t and continuation occurrence z as absent. Shape J pins the call head tick and call argument true as absent, with authored-occurrence counts guarding against a vacuous fixture. The function-body contrast is documented; paired data/function conservation is explicitly part of the repair verification, not something newly established by a paired executable control here.
The 29-atom and 13-atom populations are recorded as observations from the corrected-lexer census. The residual 302 of 3,472 is honestly labeled 'attributed by sampling, heuristic miss': the five largest sampled declarations map to existing shape H or the existing sequence-operand narrowing row by another route. This is not exhaustive occurrence classification, a new failure class, or completion of conservation. The pinned loss witnesses must be replaced by positive conservation plus discriminating red controls when their repairs land, not quietly deleted or retained as the desired behavior.
All five head-associated checks succeeded. The floor log identifies checkout e8d0c903dcfb0a7b6b2c1c39eea754b05f909f35 as the merge of this exact head into base 24ee1d006d3c95e865a0c9b28ffc0f4a4bde1b6f. Both new claims execute and pass; the accepted-drops suite reports 23/23, and the required floor reports 505 planned / 505 passed, zero failed, refused or skipped. This floor evidence is interpreted execution, not a claim of native v2 evaluation.
No new local test or full-census rerun was performed. Require the resulting merge_group candidate to pass against then-current main before landing. Approval is bound to the exact SHA above.
…into MQ-1: a fold refusal propagates as main rules; a let value and a sole statement read through body_lower_value_read; the fn-literal annotation refusal stays retired (the producer carries it)
Re-baselined census after #12313 (and #12285): exact counts, locatedInstrument: Every locus is now exact. All 3,238 Totals: authored 30,157, conserved 9,193, locus_erased 7,105, refused 488, dropped 13,371. Of the drops, 10,133 are
Located: file (rows) and lines, per shape. A match arm >=3: 495 rows in 28 files (row
|
XL-2 follow-up to #12309: two more silent drops rostered, and the residual recorded
These rows came out of the census rerun on #12313's lexer (exact loci):
v2.compiler.reference_conservation_censusreference_conservation_census_for_pathsoverreference_conservation_stratified_sample_paths, 13 batches of 25. The rerun has the same 3,472 atoms absent in accepted modules as before, and all 3,472 now match their spelling at the reported extent. No lowering change: both sites are in bright-boar-848's fold lane.New rows, each with a pinned drop claim
The claims are in
v2.test.claim.namespace_xl0.reference_conservation_accepted_drops, with subjects enrolled warm inv2.workflow.floor_pure_producer_share.match_arm_statement_body_lowers_to_its_first_expression_at_v2_body_loweringmatch_arm/stmt)body_lower_match_arm_body_capture_optional: the firstdag_surface_exprin the arm is the let's value; the binder and the continuation are droppeda_statement_after_a_let_in_a_match_arm_is_reported_dropped_holdsdata_initializer_match_reads_its_scrutinee_from_an_arm_at_v2_body_loweringdata_decl/expr/match_expr/binary_expr)body_lower_match_scrutinee_optionalsearches the whole match. On the data route the scrutinee is already folded, so the first arm's body (true) becomes the scrutinee (traced with a tree dump)a_data_initializer_match_scrutinee_is_reported_dropped_holdsLocal
gunbc run --claim-runon this head: all 8 claims in the module pass (6 from #12309 and 2 new). The rows and the roster load (--function __no_such_function__givesNoSuchFunction).Attribution of all 3,472 (rules from fixture-confirmed shapes)
an_application_binds_by_position_where_the_model_binds_by_namelist_literal_has_no_lowered_formmatch_arms_after_the_second_are_dropped_at_v2_body_loweringascast target/operandas_cast_has_no_lowered_formexpression_enclosing_a_block_headed_operand_lowers_to_that_block_at_v2_body_loweringunrecognized_primary_expression_lowers_to_its_first_atom_at_v2_body_loweringlowering_rebuilds_an_authored_atom_without_its_occurrenceResidual 302: attributed by sampling
I traced the five largest unattributed declarations by extracting each one verbatim and dumping its normalized tree. That was the bounded effort agreed with quiet-seal-543; the full tail was not chased.
v2.std.floatfloat_body_compare_ieee_ordered_outcomeAccepted { value: match a {..} }in an else branch is replaced by aniffound inside it. The source heuristic misses it.gunbc.roadmap.roadmap_site_surface_witnessstale_surface_read_does_not_ground_for_bundlelowering_accessor_collapses_a_sequence_operand, by a new route. Incall(..) && match g(..) {..}the match operand collapses to its scrutinee's first atom.test.claim.manual.outcome_diagnostics_eqwitness_diagnostics_merge_sometest.claim.runner.runner_connectivity_repair_plan_witness_testreplace_action_projects_five_step_repair_sequenceextdeps.ollama.gpu_supportollama_nvidia_driver_version_from_wire_textlet ..; match ..is replaced by theifinside the match's arm.Four of the five were under-detection of existing classes, and one was a new class, now rostered. The remaining 302 are recorded as heuristic misses, not as unexplained drops.
🤖 Generated with Claude Code