…1630); failures are sampling/quantization, not extraction (PMAT-CODE-SHIP-TWO-SECTION-68)

§67 identified 4 refinement candidates (R1-R4) for the SHIP-005
4.31pp gap. PR #1630 ships R1 (multi-block extraction) + R2
(function-targeted, prefer `def {entry_point}(` block) as the
cheapest 1-PR slice. §68 records the empirical finding from the
3-problem LIVE smoke.

Verdict: R1+R2 is a ROBUSTNESS BASELINE, not a gap-closer.

Empirical evidence (gx10 3-problem smoke, known-failed HumanEval
/1/3/6 from §67 baseline):
- HumanEval/1: FAIL → FAIL (unchanged)
- HumanEval/3: FAIL → FAIL (unchanged)
- HumanEval/6: FAIL → FAIL (unchanged)

Per-problem inspection via manual apr run: the model emits a
SINGLE fenced code block (not multiple). The block contains the
expected function but the body is non-canonical at greedy temp=0.

Failure class taxonomy:
- Class A: multi-block / wrong-block. R1+R2 fixes these.
- Class B: model-quality failure (single block, subtly wrong).
  R1+R2 cannot fix; needs R3 (Q4K → FP16) or R4 (temperature
  sampling).

The 3 sampled failures are Class B. The 4.31pp residual is
predominantly Class B; R3 or R4 are the next levers.

Methodology lesson #15 NEW: smoke-test-driven scope reduction.
A 3-problem smoke (~5 min) upper-bounds the achievable gain from
a refinement candidate BEFORE dispatching the 5h full 164-rerun.
The smoke saved a full rerun's worth of compute.

Generalises lesson #14: near-miss results need their refinements
empirically calibrated, not assumed.

Changes (1 spec file):
- docs/specifications/aprender-train/ship-two-models-spec.md
  - Atomic next action banner: v3.13.0 → v3.14.0
  - New §68 section ABOVE §63 (newest-first), 8 sub-sections
  - Cumulative methodology lessons table (#6-#15) restated

Refined R-candidate priorities post-§68:
- R1+R2: SHIPPED (PR #1630) — robustness baseline
- R3 (Q4K → FP16): NOT YET ATTEMPTED — needs FP16 safetensors
- R4 (temp=0.2 + samples=3): NOT YET ATTEMPTED — 17h gx10 compute

Spec movement:
- v3.13.0 → v3.14.0
- MODEL-1 ship %: stays at 94% (bounded path to 95% requires R3 or R4)
- MODEL-2 ship %: unchanged at 57%

Closes task #45 PMAT-CODE-SHIP-TWO-SECTION-68.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>