Dev2: pre-flight max_len check for all categories + flag stale eval r…

[Momin0000]

…esults

Objectives v3, Dev 2 (Data & Eval Integrity):

1. Ran Dev 3's gradient max_len check against every other category (diff, diff w/ multi-term, diff trig/exp/log, integrate, partial, tangent_line), not just gradient.

   scripts/verification/preflight_max_len_check.py tokenizes 2000 samples per category using the exact same path as train.py's SlangDatasetLoader._tokenize (serialize_slang_math + vocab_mapping + [BOS]/[EOS] boundaries on targets only), then reports max token length vs max_len.

   Result at current max_len=32:
   diff max_src=20 max_tgt=18 OK
   diff_multiterm max_src=20 max_tgt=18 OK
   diff_trig_exp_log max_src=23 max_tgt=23 OK
   integrate max_src=20 max_tgt=18 OK
   partial max_src=20 max_tgt=18 OK
   tangent_line max_src=21 max_tgt=21 OK
   gradient max_src=25 max_tgt=40 TRUNCATING (2000/2000)

   Confirms: gradient is the only category silently truncating today. diff/integrate/partial/tangent_line are all safely under 32 tokens and need no max_len change on their own.

   Re-ran at max_len=48 (Dev 3's proposed fix): all 7 categories, including gradient, fit with headroom (worst case 40/48). Supports Dev 3's objective 2 choice of 48 as sufficient.

   Script is reusable per Dev 3 objective 3 ("pre-flight check ... run before every future training start"): python scripts/verification/preflight_max_len_check.py [--samples N] [--max-len N].

2. docs/EVAL_RESULTS.md: flagged as stale rather than regenerated. Dev 2's regen objective depends on Dev 3's fixed checkpoint, which does not exist yet in this environment -- checkpoints/ is gitignored, and train.py's main loop currently cannot complete a run (undefined src_seq/tgt_in in the second model() call, per Dev 3 objective 1). Added a note pointing at the exact blocker and the exact regen command (python eval/run_eval.py) so the 21.7% figure isn't mistaken for current once Dev 3's other fixes land, and isn't hand-edited in the meantime.

Verified: full pytest suite run (excluding modules requiring torch/ starlette, not installed in this sandbox) -- 51 passed, 4 pre-existing failures in tests/unit/test_quotient_rule.py (quotient_rule/hessian op support, outside Dev 2 scope, unrelated to this change).

[vercel]

@Momin0000 is attempting to deploy a commit to the seno-quantum-coder's projects Team on Vercel.

A member of the Team first needs to authorize it.