Alan-Jowett / LexonGraph / 27373724964

82%

Build Type

push

github

Committed by web-flow

Commit Message

Add block-store-backed corpus sources to evaluator (#90)

* Add block-store-backed corpus sources to evaluator

Requirements: REQ-STREAM-EVAL-006, REQ-STREAM-EVAL-008, REQ-STREAM-EVAL-009, REQ-STREAM-EVAL-010, REQ-STREAM-EVAL-014, REQ-STREAM-EVAL-015, REQ-STREAM-EVAL-016, REQ-STREAM-EVAL-022, REQ-STREAM-EVAL-023, REQ-STREAM-EVAL-024, REQ-STREAM-EVAL-025, REQ-STREAM-EVAL-026.

Specs: DSG-STREAM-EVAL-005, DSG-STREAM-EVAL-007, DSG-STREAM-EVAL-008, DSG-STREAM-EVAL-008A, DSG-STREAM-EVAL-008B, DSG-STREAM-EVAL-010, DSG-STREAM-EVAL-016, DSG-STREAM-EVAL-017, DSG-STREAM-EVAL-019; VAL-STREAM-EVAL-005, VAL-STREAM-EVAL-007, VAL-STREAM-EVAL-008, VAL-STREAM-EVAL-016, VAL-STREAM-EVAL-019, VAL-STREAM-EVAL-020, VAL-STREAM-EVAL-021.

Implementation: add dual-source benchmark inputs, block-store-backed corpus loading, source-reference provenance, explicit corpus-source failure reporting, and inline/reference equivalence coverage.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Address PR review follow-up

Reject duplicate corpus source ids during profile validation, reject empty synthetic metadata keys up front, and align block-hash parse diagnostics with actual parser behavior.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Harden evaluator temp fixture roots

Use a process-local counter in addition to the timestamp when generating temporary block-store fixture roots for evaluator tests.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Validate synthetic padding block-store corpora

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Tighten evaluator corpus loading

Avoid extra cloning when chunking block-store training inputs and reject overflow when computing expected byte lengths for f32/f16 corpus embeddings.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Reuse resolved evaluator corpus inputs

Resolve block-store-backed profile inputs once per c... (continued)

Coverage Stats

542 of 726 new or added lines in 1 file covered. (74.66%)

292 existing lines in 1 file now uncovered.

9165 of 10962 relevant lines covered (83.61%)

213.47 hits per line