fix(T-2): apr finetune --max-seq-len is honoured on the wgpu instruct pipeline or refused with the documented code — never clamped to 512; one effective_max_seq_len() for every path, printed (PMAT-1009, #2924)

[noahgift]

PP-066 DAG row T-2 (spec §5 "--max-seq-len honoured or refused, never clamped"; S0-11; #2526; epic #2873; ticket PMAT-1009; receipt docs/audits/impl-PMAT-1009-receipt.md, status: partial until merged).

The defect: apr finetune --gpu-backend wgpu --max-seq-len 1024 silently trained at 512 — the wgpu instruct pipeline hardcoded 512, // max_seq_len (finetune.rs:717) and dropped the CLI value; the instruct path honoured it since #2247, so a training receipt could record a configuration the engine never ran (T-0's apr arm cannot be paired until every path is honest).

What lands: one effective_max_seq_len(requested, path) -> Result<usize> for SeqLenPath::{Instruct, Wgpu, Classify} — Some(n) is Ok(n) or ValidationFailed (exit 5, read from error.rs), None is the path's default; the 512 literal is gone; every path prints Max seq len: <effective>. Case table finetune_seq_len_truth_tests.rs (RED first at 64e3d03: the function did not exist): {256, 512, 1024, 2048} × three paths, the default row, the refusal-code row. Contract contracts/apr-finetune-config-truth-v1.yaml (kind: pattern); README 1812 contracts.

Acceptance, re-run by the orchestrator on 2beca0d

A_i	rc
cargo test -p apr-cli --lib finetune_seq_len_truth	0 (4 passed)
cargo test -p apr-cli --lib finetune (80 tests)	0
cargo fmt --all -- --check · cargo clippy -p apr-cli --lib -- -D warnings	0 · 0
pv validate · pv lint · check_contract_test_binding.sh · check_contract_enforcement.sh · check_readme_claims.sh · check_no_claim_literals.sh · check_roadmap_diff_additive.sh	valid · PASS · 0 ×5

Mutation — RED, then restored GREEN: the Wgpu branch returns Ok(512) regardless of the request → effective_max_seq_len_wgpu_never_clamps_to_the_old_512_literal FAILED, effective_max_seq_len_honours_every_requested_value_on_every_path FAILED (2/4) → restored 4/4.

Recorded: the card's --test finetune_seq_len_truth integration target became a #[cfg(test)] module of commands::finetune (an integration test cannot reach the crate-private commands tree; the lib.rs seam is blocked by the pre-commit complexity gate on pre-existing debt — same finding as R-3); the test drives the pure function, an end-to-end wgpu run is T-0's harness. Receipt for this PR itself: advisory, not produced (driver A1).

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3005 head=ea91bbb96507e64b0640ba20982b98f930ad351a 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.