[PhysShell]

What this is

Evidence for the P-014 Tiny PoC, under the owner's decision GO-P014-TINY-POC (after #421). It contains only
documentation, a research instrument, synthetic fixtures and data, all under
docs/evidence/p014-tiny-poc-2026-10/.

Result: PASS-P014-TINY-POC. It covers a research instrument that measures binding. It does not allow
production work.

• Nothing changes in the extractor, the core, OwnIR, goldens or CI.
• No findings were produced or changed.
• Extractor: project-aware semantic resolution — types from SDK implicit usings do not resolve #407 stays deferred, and docs(evidence): P-014 — record owner ruling (GO-P014-BOUNDED-DESIGN / NO-PRODUCTION) #420 is not touched.

Owner decision on this evidence: ACCEPT-POC-EVIDENCE / GO-PR / NO-PRODUCTION. The next production stage needs a
separate decision.

Order of work (committed history)

1. 5a9bb32 — pre-registration (criteria, KILL, time and memory limits, the SDK conflict contract), committed
   before the instrument existed.
2. 207ef84 — Amendment 1, committed before any run. The docs(evidence): P-014 reference coverage — GO-P014-BOUNDED-DESIGN / NO-PRODUCTION (research only) #419 per-site lists had never been committed, so the
   comparison changed to sites resolved in legacy and not in exact.
3. 35e2c8c — instrument, fixtures, controls against the SDK oracle.
4. cb9519d — real projects P4/P5, cost, report.
5. a685402 — the four limits of the PASS, written out explicitly.

Results

Area	Result
Project isolation	0 foreign bindings across all controls, P4 (4 projects) and P5. K2, K4 (CS0101 in one shared compilation, CS0433 at the site when projects are separate), F2.3–F2.5 and the guard self-test all pass
Reference selection	K8a, K8b, K8c, K9 and K3 manifests equal the SDK's own resolved references on my synthetic fixtures, restored offline from the global packages folder. The "framework wins on a tie" assumption was never needed
Coverage sidecar	deterministic: byte-identical in 3 of 3 runs for every mode and project; invariants hold
Calls resolved, exact vs legacy	P4 62.0% vs 25.1%; P5 64.1% vs 32.1%
Cost, P5	warm run 1.43× legacy (limit 3×); peak about 208 MB
OwnIR and diagnostics	unchanged; today's extractor facts are byte-identical across 3 runs

Calls resolved in legacy and not in exact. P5 has none. P4 has 22, and all were read.

• Cause. Implicit usings are not applied in the exact mode, as the owner required. In the separate per-project
  compilation, the error types they leave behind show up at the call site.
• Legacy over-counts. It counted these calls as resolved although their signatures contain error types.
• No wrong binding. None of the 22 is a wrong binding in the exact mode.

#419 R3 sites. All 7 bind to real IEventSymbols, including the actionable ObservableObject.PropertyChanged.
The 6 FPs are a problem of the lifetime model, not of references.

Limits (REPORT §8)

1. References are not the whole source set. A COMPLETE reference set does not mean a complete compilation:
   P4 and P5 still have an incomplete source set (implicit usings, generators, XAML).
2. Conflict rule qualified only in part. It is shown only on the forms checked here:
   • package vs package with the same DLL name is not qualified;
   • the rank and platform steps are not qualified;
   • the full SDK contract is not qualified.
3. The host's .NET changed during the work. The 8.0.31 and 10.0.12 packs are gone, so the old measurement
   cannot be reproduced on its original packs.
4. P5 picks up a Directory.Build.targets from outside the repository. It is reported only; there is no
   guarantee about configuration that comes from outside the repository.

Disclosed deviations

• Three defects in my own instrument and comparison script were fixed before the real-project runs:
  • a type inside new T();
  • the order the K1 check expected;
  • the package version, which the oracle gives only in FusionName.
• One fixture run used a stale binary and was discarded.
• The read-only diag mode was added after the measurements.

Checks

• Diff scope. The diff touches only docs/evidence/p014-tiny-poc-2026-10/; git diff --name-only 228e38d7
  lists nothing outside it.
• Secrets and paths. ruff check . passes; gitleaks found nothing on the 5 commits; there are no local paths
  (they are replaced by <nuget>, <dotnet>, <research> and <work>).

🤖 Generated with Claude Code

https://claude.ai/code/session_01TFkNAFFQz9KoTzpPws8vYG