trainlog/docs/reviews/training_knowledge_v1_engineering_review.md

Training Knowledge V1 — engineering checkpoint

Date: 2026-09-09. Status: TRAINING_KNOWLEDGE_V1_ENGINEERING_REVIEW=PASS.

Scientific review separately passed within its documented scope and uncertainty. The independent temporal delta review returned TEMPORAL_DELTA_REVIEW=PASS with no findings. One configured final-reviewer completed an initial full-tranche audit, reported the findings recorded below, and its authorized bounded repair chain completed. Independent bounded verification returned FINAL_REVIEW_REPAIR_VERIFICATION=PASS; fresh final executable validation also passed. TRAINING_KNOWLEDGE_V1=PASS.

Current implementation and contract

Six canonical scientific catalogs feed generated C data and the Android asset loader. Both clients expose scientific lookup/filter APIs and runtime context composition. Android has a collapsible knowledge section; the TUI has a UTF-8 cell-aware scrolling knowledge screen tested at 72×20. No schema, scientific catalog, frozen exchange artifact or stored user history changed in this resume.

The temporal repair covers the two new C readers and their Android counterparts: occurrence pagination and latest explicit MAX. The established operational profile and parser-only extension classification are recorded in the temporal contract. All use exact integer-second/fraction comparisons and bytewise stable-ID ties. Pagination scans all matching metadata, retains at most limit + 1, then hydrates only selected rows under one read snapshot. Emitted cursors preserve source text.

C has the existing 40-character timestamp output limit: long valid candidates are compared exactly, and a selected unrepresentable result fails explicitly. Invalid caller cursors and malformed stored timestamps have distinct error paths. No malformed date can disappear through SQLite NULL comparison.

Python validation uses the same explicit grammar and exact chronology. The parent reproduced and repaired an optional-checker dependency issue after the worker implementation: a strict JSON Schema checker could reject omitted seconds before Trainlog validation. A private per-call override now aligns the schema-format check while preserving all other checker configuration. Actual validation-path tests cover absent, strict and permissive optional checkers.

Original defect retained as historical evidence

The original implementation combined an offset-limited cursor guard with SQLite julianday() and a different Android parser. Its production probe was:

Newer source timestamp	Original outcome	Current independent probe
2026-09-10T10:00:00+02:00	control passed	PASS
2026-09-10T10:00:00+14:30	emitted cursor rejected	PASS
2026-09-10T10:00:00+15:00	newer occurrence omitted	PASS
2026-09-10t10:00:00Z	newer occurrence omitted	PASS

The independent current probe extends this to twelve spellings. Every case uses two sessions in a temporary database, page size one, exact source/cursor assertions, continuation and exhaustion. Both normal and sanitized builds pass. Original evidence remains /tmp/trainlog-knowledge-validation/cursor-probe.c and cursor-probe.log; fresh final probe evidence is under /tmp/trainlog-knowledge-review-resume/.

Final validation evidence

Check	Result
meson compile -C build	PASS, fresh strict warning-as-error build
meson test -C build --print-errorlogs	42 passed
Knowledge validator/generation tests	PASS; eight Python tests
Temporal Python tests	PASS; four tests including optional-checker independence
JSON/import-contract validators	PASS; six import cases
Three standalone public C17 headers	PASS, including -pedantic-errors
Two targeted C tests under ASan/UBSan	PASS; timestamp Python test also passes in that test invocation
Independent twelve-form C production probe	PASS normally and under ASan/UBSan
Android unit tests and debug assembly, Java 17	PASS; 56 tests, zero failures/errors, one known fixture skip
Project skill validator and git diff --check	PASS

The known Android skip is RealAndroidV9BodyZonesMigrationTest, because its external TRAINLOG_ANDROID_V9_FIXTURE is unavailable. No knowledge/temporal test is skipped. Manual MTP, installed-device UI and real-terminal visual checks were not performed or claimed. Sanitizer evidence is scoped to the new affected C paths, not the complete GUI executable.

Independent review and final-review repair record

The isolated temporal review was independent of the original implementation and reviewed grammar/calendar constraints, offset bounds, fractions, byte ties, cursor aliasing/exclusivity, source capacity, snapshots, and Android/Python parity. It ran Meson training_context and timestamp_validation plus four Python temporal tests, returned TEMPORAL_DELTA_REVIEW=PASS, and required no repair.

The configured final-reviewer then audited the full tranche. Its initial result was FAIL and identified these concrete findings:

Severity	Finding	Authorized repair recorded in /tmp/trainlog-knowledge-review-resume/repair.md
BLOCKER	Stale temporal-defect documentation in architecture.md, exercise_data_model.md, and tui.md	Replaced claims of an unresolved reader defect with the independently passing temporal contract and correct repair/validation state.
BLOCKER	Android loader parity holes	Enforced canonical exercise/equipment identity syntax, reverse capability compatibility, HIGH evidence source type, and non-unresolved BODY ZONE audit evidence; added loader regressions.
BLOCKER	Meson generator inputs omitted body-zones-v1.json and equipment-v1.json although validation reads them	Declared both files as training_knowledge_generated inputs and proved regeneration without output-byte change.
HIGH	C query accepted a muscle role without a muscle ID, unlike Android's paired filter	Rejected role-only queries, documented the paired contract in the public header, and added a C regression.

The repair report records strict C17 syntax, knowledge validation, generated output equality, targeted Meson test, Android loader tests, dependency proof, and git diff --check as PASS. Independent bounded verification closed all four findings and explicitly returned FINAL_REVIEW_REPAIR_VERIFICATION=PASS and TRAINING_KNOWLEDGE_V1_ENGINEERING_REVIEW=PASS. The fresh tester matrix then passed: strict build; 42 Meson tests; eight knowledge and four temporal Python tests; knowledge/JSON/import validators; three strict C17 headers; affected C knowledge/context tests under ASan/UBSan plus timestamp validation; normal and sanitized 12-form temporal probes; Android 56 tests with zero failures/errors and one known missing-fixture skip; Java 17 debug assembly; skill validation; and diff check. Fresh generated C is byte-identical with SHA-256 e8c099f67eb111d61621b5d76592c049823af5508d43e73ec646f22e4c377fca; all six Android assets are byte-identical.

Remaining observations

The prior bounded review's secondary scientific BODY ZONE IDs on the TUI knowledge page remain a nonblocking presentation inconsistency. Local machine model/execution uncertainty and missing real exercise UUIDs remain as recorded in the scientific review; no phantom identity or scientific reassessment was introduced. Older lexical timestamp readers, arbitrary-length C output and verified leap-event support are outside this bounded repair.