DeltaWire: measured performance envelope + first real-model admission run

[bigboateng]

Executes the approved Iteration plan (labs/20-deltawire/docs/13-performance-and-admission-plan.md): move DeltaWire from estimate-only selection to a measured performance envelope (Iteration 1), then run the first admissible real-model B0/D1 comparison (Iteration 2). All prior conformance guarantees (determinism, runtime-boundary, zero-model core) are preserved.

Iteration 1 — performance & conformance closure

• deltawire bench (hidden subcommand, internal/cli/bench*.go): runs the real pipeline per family × scale, emitting compiled/output bytes + phase timings. Determinism and runtime-boundary preserved. Profile collected in eval/results/perf-profile-v1/.
• optimizer.PerfProfile + AnalyzeWithProfile (internal/optimizer/profile.go): optional measured coefficients threaded through candidates/select, with safe fallback (profile_missing / _stale / _missing_family reason codes). Analyze == AnalyzeWithProfile(…, nil), so no caller churn.
• Closes GAPS.md item 10 (Pareto report carries measured bytes/timing).

Iteration 2 — real-model admission pack + first call

• entitlement-matrix task authored under eval/admission/ (deliberately not eval/tasks/, which is a closed generic-oracle system) with a frozen independent oracle. Deterministic admission gates verified by internal/optimizer/entitlement_test.go.

• eval/scripts/admission_runner.py: fail-closed, content-addressed two-arm runner scoped to the admission task, reusing the Harbor invocation + fidelity receipts. --dry-run validates every hash with no model call.

• First B0/D1 pair (gemini-3.1-pro-preview, Harbor 0.20.0), initial construction:

  arm	DeltaWire offered	used	oracle exact	end-to-end	input tok	wall
  B0 (control)	no	n/a	1	pass	39,949	55s
  D1 (treatment)	yes (--skill)	no	1	fail*	68,585	81s

  * Not a task failure — D1's artifact is correct and byte-identical to B0's. The treatment agent never adopted the framework (classification=not_used, plan_path=null); it wrote plain Python like the control, for ~29k more input tokens. This lands on the interface/adoption branch of the plan's decision rule. Full write-up + follow-ups in eval/results/admission/entitlement-matrix/README.md. Raw Harbor evidence retained.

Key finding

At this scale (54 records, one-shot-correct in plain Python) DeltaWire is neither chosen nor needed when offered optionally. Honest next steps: (1) raise scenario stakes to where the framework's determinism/trust-surface actually bite; (2) make the treatment arm require the framework to measure interface usability at all. This motivates a deterministic trigger mechanism (cf. the Pitot integration) rather than relying on model adoption. The 2.3 modification arm is not yet run.

Non-goals held

No universal performance claim; no supervisor-contract change; no model call until every deterministic admission gate was green.

Verification

• go build ./... + go test ./... green.
• scripts/validate.sh EXIT=0 (determinism, runtime-boundary, production-cleanliness; model_calls: 0 on the deterministic path).
• Live API key kept out of the repo (gitignored eval/.secrets/, scanned before commit); staged linux binary gitignored under eval/admission/*/environment/.generated/.