docs(evidence): #400 dictionary-ref PoC — KILL, archived evidence

[PhysShell]

Summary

This PR archives the bounded Roslyn PoC for the #400 dictionary-ref pilot. Its result is
KILL-DICTIONARY-REF-PILOT. Owner ruling: ACCEPT KILL / ARCHIVE AS SEPARATE RESEARCH PR, outcome
STOP-PRODUCTION / KEEP-EVIDENCE.

The pilot was authorised by the decision GO-400-DICTIONARY-PILOT. That decision and the research report are in
#412.

The PR adds evidence only, under docs/evidence/dictionary-ref-poc-2026-10/. It does not change production code,
OwnIR, CI or goldens, and no implementation PR follows from it.

Result

• Targets. Both buggy revisions warn at the right lines:

  • Meziantou DictionaryExtensions.cs:31 and :56;
  • MAUI BindableObject.cs:793, through CreateContext → defaultValueCreator(this).

  Both fixed revisions are clean. For MAUI this is trivial, because the fix removes the API from that method.

• Controls. The mandatory negative controls are silent: no mutation, a different dictionary, a safe factory,
  and a ref used only before the mutation.

• Precision gate (≥ 80 %): FAIL, as measured, with no retroactive adjustment:

  Round	Probe	Warnings	TP / FP / UNCLEAR	Precision	Conservative
  1	v1	18	9 / 5 / 4	64 %	50 %
  2	v2	14	8 / 3 / 3	73 %	57 %

• Severity of the 17 sample TPs. 0 were proven at runtime. 3 are contract violations without observable effect;
  14 are plausible but not demonstrated. Runtime-proven defects exist only for the two targets (docs(evidence): next real-bug class selection — #400 dictionary-ref pilot (#382/#406 NO-GO) #412).

• Third-party analyzers: NOT_TESTED. The package download was not authorised; this is not a MISS.

• Not authorised. The post-hoc ref = factory(…) observation is recorded as not evidence. No third round is
  authorised.

Contents

• REPORT.md, which covers the gates, precision, the effect split, the limitations and how to reproduce.
• PRECISION-PREREG.md, verbatim.
• Frozen probes probe-v1/ and probe-v2/, checked by SHA256SUMS.
• controls/.
• data/:
  • sample lists pinned by blob SHA;
  • every warning with its label and reason;
  • outputs for the targets and the controls.
• scripts/ for search, index derivation, fetch, blob check and the full re-run.

No third-party source is committed: samples/ and targets/ are fetched by SHA and gitignored.

Verification

• scripts/verify_blobs.sh: r1 429/429 and r2 252/252 files match their blob SHA.
• scripts/make_index.py --check: both sample indexes are re-derived byte-identically.
• scripts/run.sh regenerates data/ and prints the two tally lines above.
• ruff check, ruff format --check and mypy --strict pass on scripts/.

🤖 Generated with Claude Code