You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
(OPT) Retain independently derived renewal directories across ordinary construction and replay #783
Current source discovery at 5c39cad found repeated whole-function renewal reconstruction: hir::vec_loop_renewal::binding_in calls bindings_in(...).remove(at), and bindings_in collects all ordinary + consuming sites. PlanBuilder::renewal_binding holds only byte-helper sites and takes this fresh route; ordinary cleanup replay asks renewal_profile::binding_with_index again for the same site. SkeletonWork already independently owns a renewal directory. This is distinct from the completed Bytes helper ModuleIndex.
Retain one ordinary RenewalBindings directory per actual construction/replay phase, independently derived from exact current program/function pointers. Preserve family selection priority and per-family membership: schema gating cannot use union membership alone. Selected primitive receiver authority remains separate. Reuse directory lookups without rescanning/rebuilding maps per occurrence; precharge actual held backing/work before allocation and preserve sticky first refusal and Drop.
Acceptance: saved old per-query classification vs phase-held directory across ordinary/owned/consuming families and opposing identity orders; malformed/stale/source-pointer/family-schema controls; exact/one-short/no-refund/Drop for actual constructor and replay; unchanged cleanup inventory/order and whole original application controls. No guessed quota/forecast discount or skipped validation. This is real repeated work but is NOT evidence identifying the earlier HirLowering 10280-byte refusal, which occurs in another phase. No application-fit or live agent token advantage claimed. Implement then minimum owning checks in the large release-branch batch.
Implemented and qualified the bounded phase-held renewal-directory contract. Construction and replay derive separate exact-program/function directories, retain independent family membership and original selection priority, and pay real backing/work before acquisition. Selected primitive authority remains separate.
Published acceptance audit and complete raw provenance include the previously unpublished fd2ac seven-body helper-index control: all seven passed, including copied/foreign source, changed-provider, old-schema, exact/short/zero, spare-backing and sticky-publication cases. Its raw build receipt, metadata and runtime before/after executable hashes agree. The audit binds the later actual family/constructor/replay/presence controls, preserves the earlier phase failure, and verifies 33 complete relevant files against their qualified predecessors. This is historical multi-subject evidence, not a whole-current-head PASS claim.
Existing codec, ShiftSim, Catalog and normalizer application failures remain explicit. This closes the held-directory request; it does not claim application fit, recovery of the original HirLowering refusal, or measured agent-token savings. No limits, schemas, cleanup order or tests were weakened.
Activity
wsdt commentedon Oct 16, 2026
Implemented and qualified the bounded phase-held renewal-directory contract. Construction and replay derive separate exact-program/function directories, retain independent family membership and original selection priority, and pay real backing/work before acquisition. Selected primitive authority remains separate.
Published acceptance audit and complete raw provenance include the previously unpublished fd2ac seven-body helper-index control: all seven passed, including copied/foreign source, changed-provider, old-schema, exact/short/zero, spare-backing and sticky-publication cases. Its raw build receipt, metadata and runtime before/after executable hashes agree. The audit binds the later actual family/constructor/replay/presence controls, preserves the earlier phase failure, and verifies 33 complete relevant files against their qualified predecessors. This is historical multi-subject evidence, not a whole-current-head PASS claim.
Existing codec, ShiftSim, Catalog and normalizer application failures remain explicit. This closes the held-directory request; it does not claim application fit, recovery of the original HirLowering refusal, or measured agent-token savings. No limits, schemas, cleanup order or tests were weakened.