# Suite35 native AQ integration revision

The default run must use the established AQ project directly: `./run_suite35.sh`. No hand-built grades CSV, calendar or mapping config is required.

Source-backed contracts recovered from the uploaded suites:
- `train_frame`/`test_frame` contain `IDSV`, `IDMH`, `target_NHHK`, usually `IDKQDK`, and numeric `target_grade10`. Canonical `Xtr/Xte` and `ytr/yte` are full-frame aligned. Historical numeric grades already exist; binary labels are never reverse-mapped to invented numeric grades.
- Shared raw columns are intersected across requested faculties and folds. The default adds the known 17 course fields. `course_target` is a numeric feature matrix, not a course-ID vector.
- Sidecar `train_row_index` maps to finite CF-covered canonical rows, including finite-probability checks. Target order is canonical unless an explicit target index is present. Existing aligned values are retained. Missing/uncovered features can be reconstructed from own static metadata plus strict prior-semester event statistics.
- Semester order is `year*3+term-1`, term in 1,2,3. No release calendar is present. This run adopts the established previous-semester-grade availability assumption, labeled `semester_order`; ordinal carriers exist only to reuse internal comparison code and are never represented as observed publication dates.
- Full historical train frames form the deduplicated SSL event bank, independently of old CF coverage. Target-frame grade values do not enter that bank. Grades reused in later training caches remain constrained by each prediction/SSL semester cutoff.
- Default downstream support retains all canonical labeled training rows as in Suite26; a covered-support option is available for legacy comparisons. All current Suite35 arms share support. The audit reports the population change from Suite25/32 covered-only fitting.
- Loading is lazy by faculty/fold. Shared event metadata can be retained globally; all 98 raw feature matrices must not be retained simultaneously. History tensors use compact padding and retain a unique-prefix bank plus row map. Frozen/evaluation prefixes are encoded once before mapping compact embeddings back to enrollment rows; trainable encoders expand only minibatches, preserving enrollment weighting.
- Existing canonical and structured caches are never modified. Split CSV fallback follows the known expanding-fold layout and target semantics; any new intermediate output belongs under Suite35's own output directory.

Implementation tasks:
1. Native schema loader/discovery, source and chronology audit, existing sidecars and safe reconstruction, split fallback, lazy cohorts, source fingerprints.
2. Integrate `semester_order` into the data contract; automatic native default CLI and single-log shell; maintain explicit JSON demo compatibility.
3. Retain all three encoders/SSL objectives/transfer controls and the existing chronological selection/freeze protocol. Bound host memory and duplicate-prefix preprocessing.
4. Verify with fixtures in the actual previous-suite schema (not pre-normalized CSVs), no-input-file CLI invocation, sidecar reconstruction/fallback, source mutation rejection, native development+retrospective training/replay, existing unit suite and focused review.
5. Replace the delivered package and first-run documentation with verified single-command instructions. State that server execution and real AQ scores remain unverified locally.
