test: fold train loader try_load_apr coverage

[noahgift]

Folds test/cov-train-next (06f0f38). Formals use declared Σ glyphs; nt, witness and README count regenerated. README count may conflict with other folds that add a contract: merge main after they land.

keep-open: test-coverage tickets stay open until their owners confirm
ont-delta: none test-only fold, no ontology delta

🤖 Generated with Claude Code

[noahgift]

seat N, aprender-triage-worker @d234c5a0b53166e6cc6c8493de12032d6dcfb8b6 APPROVE

Read: tests-only fold — 4 try_load_apr tests in aprender-train config/train/loader/tests.rs (missing/corrupt/empty dir -> None, metadata checkpoint_step beats file-name step, directory picks newest model-step-N.apr, weight round trip with a training tensor present), the train-loader-try-load-apr-v1 contract (Σ formals) and regenerated README/nt/census/witness. No production-code change seen. Not run: tests, mutation, CI. Own trailers: 0.

[noahgift]

seat 2, aprender-pace @d234c5a0b53166e6cc6c8493de12032d6dcfb8b6

Read: tests-only, no production code changed. 4 new tests in loader/tests.rs cover try_load_apr_for_inference: a missing path, a corrupt file and an empty dir all give None; metadata checkpoint_step beats the file-name step; a dir picks the newest model-step-N; and weights round-trip with a training tensor present. The checkpoint_step metadata read sits at data.rs:165. The new train-loader-try-load-apr-v1 contract passes pv validate, and its formal lines use only the declared Σ form. README goes 1983 to 1984 (+1 contract), contracts.nt gets 5 triples, and the ONT witness json is regenerated. The tests use expect(), not unwrap(). 0 aprender-pace trailers. I did not run the tests.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5592 head=d0987c48d3b1a8c880cae8bfaa72dc610af9603c verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.

[noahgift]

seat 1, aprender-triage-worker @c889607d2c7096b7886cadf01974af3ebb429cb3 APPROVE

Re-seat. Own-trailer count 0. Since my seat at d234c5a the PR gained e93f362 (4 try_load_apr tests via try_load_apr_for_inference in aprender-train loader/tests.rs, +75, and new contract train-loader-try-load-apr-v1.yaml) and a merge of origin/main (witness regenerated).
Read: the four tests (missing/corrupt/empty-dir -> None; metadata step 42 beats file-name step 5; directory picks model-step-12 over -3; weights round-trip with a training tensor present) and the contract. Assertions are specific.
Note: the contract lists 2 invariants but obligations and the single FALSIFY entry cover only the metadata and directory-step tests; the weights round-trip invariant has no obligation id. Not blocking.
Not run: tests, pv validate, merge-from-main content. This is one seat; a second reviewer is still needed.

[noahgift]

seat 1, aprender-pace @c889607d2c7096b7886cadf01974af3ebb429cb3

Read the full diff (+121/-5, 5 files). Only code change is four tests in aprender-train loader/tests.rs (missing/corrupt path, metadata step over filename step, newest step in a directory, weight round-trip with a training tensor present); no production code touched. I did not check the asserted step semantics against try_load_apr_for_inference's source, and I did not run them. README/contracts.nt/witness are generated; not regenerated by me. 0 aprender-pace trailers. CI not read.

[noahgift]

seat 1, aprender-triage-worker @d5d93cb1f165eb915ca40f34bce397cec67cb714 APPROVE

Re-seat. Own-trailer check: 0. My last seat was @d234c5a0b5. Restricted to the PR's file set, the delta is 4 files, all derived from the merge of main:

• README.md: the two CONTRACT_COUNT markers 1984 -> 1985.
• contracts/contracts.nt: 7 triples for the new course-flow-batch-wip-v1 contract (from main), plus one referencedPath swap ("crates/aprender-compute" out, "docs/ssc-training-status-2026-03-22.md" in).
• contracts/shapes.ttl: 74 lines, regenerated (paired -/+ lines; I did not trace each one).
• the witness json: census_id_set_sha256 and reasoner_git_sha refreshed; the reasoner sha is d234c5a, my last-seat head, not the merged head, which is normal for a regen but is the one thing a reader may query.
  Not run: no regeneration, no tests. CI was still pending (4 pending, 2 skipped, 1 success) and merge state BLOCKED, so I read no green. I arm nothing.

[noahgift]

seat 1, aprender-pace @d5d93cb1f165eb915ca40f34bce397cec67cb714

Re-seat after a merge of main plus a shapes.ttl regen. Delta from my previous seat @c889607d: on the PR's own diff (generated witness/contracts.nt/README excluded) only contracts/shapes.ttl moved (34+/34-, per-slot sh:node IRIs); the PR's content (the train-loader-try-load-apr contract and its tests) is unchanged. 0 aprender-pace trailers.

Evidence on the regen is a3's, not mine: a3 reports running ont4b_shapes_gate::the_tracked_repo_graph_is_fresh from this tree after a forced pv rebuild (1 passed) and that a second extract gives no diff. I did not run it. One thing I could not explain from the diff: this head's shapes.rs is identical to origin/main's, yet origin/main's own shapes.ttl still has the old (non-per-slot) IRIs, so either main's committed file is stale against main's generator or the PR's new contract changes how slots are named; I only compared the two files, so I leave that to CI. CI had not finished when I looked (shards and x86-main queued), so this is a content seat, not a green-checks statement.