research: full research programme + pre-registered evaluation refuting two reliability mechanisms

[CodeWithJuber]

Summary

Uploads the full research programme behind forgekit as a single research/ tree with a master
index, and adds the pre-registered empirical evaluation that refuted two of the programme's own
headline claims.

That refutation is the contribution. The later work corrects the earlier work, and the corrections
are the most useful part — so research/README.md leads with what was actually measured, not with
what was originally claimed.

What the measurement showed

	Claimed (self-built demos)	Measured (real data)
Impact oracle recall	1.00	0.022 — 95% CI [0.012, 0.033]
Router/gate F1	1.00	0.37 [0.12, 0.59]
Cost saving	+62.1%	−20.2% (total pipeline spend)

91% of files got exactly zero recall, and a grep baseline with no dependency graph beat the
graph-based oracle by ~10× on F1. The cost result is the sharpest: routing genuinely saves 59.5% on
first attempts, but almost none of that cheaper output is correct, and once escalation retries are
counted the pipeline costs more than always using the premium tier. The saving did not shrink —
it inverted.

Diagnosis, then repair

Classifying all 18,221 false negatives showed 96.8% were not a ground-truth ceiling: the path
existed in the graph the tool already builds. 94.7% were sibling pairs reachable only by a
mixed-direction path that reverse-only traversal cannot take, plus a construction defect that broke
one of the two standard Python package layouts.

Repairing both, with all eight new parameters fixed on six repositories and frozen before the three
held-out repositories were touched even once:

	Precision	Recall	F1
As shipped	0.398	0.022	0.042
+ construction fix	0.304	0.242	0.270
+ traversal fix, held-out	0.305	0.653	0.416
grep baseline, held-out	0.269	0.601	0.371

Beats the baseline for the first time. The margin is narrow, and precision generalised distinctly
worse than recall (−0.141 vs −0.021) — visible only because the split was declared in advance.

What is in this PR

research/
  README.md                  master index — leads with the refutation
  cognitive-substrate/       the theory: 48pp whitepaper, literature gap map,
                             graded evidence layer, ecosystem map, the fourteen-mapping
                             design lens, primary sources, figures
  formal-synthesis/          (already in repo) the mathematics
  empirical-refutation/      the pre-registered evaluation, repair, and replication package
  python-prototypes/         (already in repo) the code

Honesty notes a reviewer should check

• EXECUTIVE_SUMMARY.md carries a status banner saying its own prototype numbers are refuted,
  with a table of what replaced them. Uploading it unmarked would have been misleading.
• Three widely-repeated industry statistics were caught as misattributed during evidence grading
  and dropped rather than repeated.
• Priority is conceded. The two-layer composition law is standard protection-layer algebra, and
  two concurrent preprints derive a strictly more general form weeks earlier. What survives is that
  both are simulation-only; this work has real repositories.
• A theoretical claim is corrected. Perfect recall was inferred from a completeness theorem, but
  such a theorem guarantees completeness only relative to the relation the closure runs over.
• Our own adversarial review ships with the work, listing four numbers we had stated incorrectly
  before correction — including two cases of reporting the more favourable of several available
  figures, which is exactly the failure the paper indicts.
• The correctness judge is the same model as the ladder's mid tier, so for 38/64 tasks it grades
  its own model family. That biases toward overstating savings, which makes the reported inversion
  conservative.
• The freeze was self-administered with no external timestamping authority. A reader can verify
  internal consistency and the append-only amendment trail, but must take the ordering on trust.
• One figure (eval_precision_recall.png) had a panel title claiming the oracle held the highest F1
  while the same panel plotted grep higher. Corrected and marked superseded.

Review notes

• No source code changed. The repaired oracle ships inside the replication package rather than
  replacing research/python-prototypes/impact_oracle/, so that swap stays your decision.
• Opened for review, not for immediate merge.