Skip to content

Recover runtime candidates with bounded reference and caller evidence - #1770

Merged
gmh5225 merged 13 commits into
devfrom
codex/runtime-context-evidence
Oct 25, 2026
Merged

gmh5225 merged 13 commits into
devfrom
codex/runtime-context-evidence

Conversation

@gmh5225

@gmh5225 gmh5225 commented Oct 25, 2026 •

Copy link
Copy Markdown
Contributor

Short or byte-identical runtime routines remain unresolved when the distinguishing information is an import slot, a vftable type, a function-address relocation, or an established caller. Add bounded validation of those constraints and complete-body incoming records, with raw cross-version collision checks in the generation tools. Incoming records use a name index, stay out of global byte scanning, and extend only from a strong rooted frontier. Candidate evidence never authorizes automatic source names.

Also validate cached context independently of parallel name materialization, reuse import-veneer resolution, and stream the existing candidate JSON format. Preserve exact alternatives, reject unsupported reference forms, and expose incoming evidence through the SDK and GUI.

Validation:

  • 220 native signature tests and 5 SDK tests passed in the existing LLVM 23 r6 validation build; 114 focused tests passed with the changed translation units instrumented by ASan/UBSan.
  • 73 Python generation/refinement tests passed. The three affected core/SDK translation units also compile against the rebased development headers.
  • 23 stripped MAP-backed probes with the companion data: 1,365 additional correct candidate addresses, no newly wrong candidates or lost correct results, and no automatic names. The tracked primary gap set improves from 4/161 to 134/161. Existing candidate errors remain; this is not a universal accuracy claim.
  • Latest development changes for original-byte matching and analysis snapshots are retained. The repository documentation check currently reports an unchanged upstream section-marker error in docs/process-emulation.md.

The companion data PR records compiler-library provenance and hashes. Performance measurements and raw probe artifacts remain local; full-process measurements are noisy on the busy local host.

Incremental incoming settlement preserves all 23 probe JSON outputs exactly. Required-context selection retains an independently distinguished claim even when unrelated same-byte siblings remain ambiguous, without promoting optional references or incomplete alias evidence.

Pins signatures to merged revision 39c99bbade0bf21c00cc0d70df7c577d98e34a7c from NeverSight/signatures#25.

@gmh5225
gmh5225 merged commit 879c953 into dev Oct 25, 2026
6 of 29 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants