@@ -2894,6 +2894,120 @@ jobs:
28942894 [ "$rc" -ne 0 ] || { echo "FAIL: a broken candidate exited 0"; exit 1; }
28952895 echo "OK: a broken candidate is a visible failure (exit $rc), never a Python rescue"
28962896
2897+ # #263-A — the performance measurement INSTRUMENT, calibrated. Not a
2898+ # baseline: this job proves the instrument stands up, on both platforms, and
2899+ # that it refuses to measure anything decisive before the D7 freeze.
2900+ #
2901+ # It publishes no engine comparison and can publish none: the harness holds
2902+ # one engine's samples per cell and refuses a report carrying a comparison
2903+ # key. What it asserts here is that the refusal is real, not that Rust is
2904+ # anything.
2905+ perf-instrument :
2906+ name : " #263-A measurement instrument (CALIBRATION_ONLY)"
2907+ strategy :
2908+ fail-fast : false
2909+ matrix :
2910+ os : [ubuntu-latest, windows-latest]
2911+ runs-on : ${{ matrix.os }}
2912+ defaults :
2913+ run :
2914+ shell : bash
2915+ steps :
2916+ - uses : actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
2917+ with :
2918+ # The step 4 freeze controls read git objects at the frozen source
2919+ # commit and prove it is an ancestor of this head. A depth-1 checkout
2920+ # cannot answer either question, and those controls refuse rather than
2921+ # pass vacuously — so the history is the thing that makes them real.
2922+ fetch-depth : 0
2923+ - uses : actions/setup-dotnet@67a3573c9a986a3f9c594539f4ab511d57bb3ce9 # v4
2924+ with :
2925+ dotnet-version : " 8.0.x"
2926+ - uses : actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
2927+ with :
2928+ python-version : " 3.13"
2929+ - uses : dtolnay/rust-toolchain@fa04a1451ff1842e2626ccb99004d0195b455a88 # master, 2026-07-10
2930+ with :
2931+ toolchain : stable
2932+ - name : Build the production own-cli candidate
2933+ working-directory : rust
2934+ run : cargo build -p own-cli --release
2935+ - name : The instrument's own selftest (no measurement taken)
2936+ run : python scripts/perf_baseline.py --selftest
2937+ - name : The instrument controls (firewall, both identity domains, provenance)
2938+ run : python tests/test_perf_instrument.py
2939+ # The Round 7 apparatus is a classifier and a structural preflight, both
2940+ # exercisable without a clock. The exclusivity control is the one that
2941+ # matters: the preregistration's first outcome table let one dataset
2942+ # satisfy two contradictory verdicts, and nothing in the process that
2943+ # produced it would have noticed.
2944+ - name : Round 7 apparatus controls (outcome exclusivity, zero guard, B1-B4)
2945+ run : python tests/test_round7_apparatus.py
2946+ # The calibration policy is pure: no clock, no files, no randomness, and no
2947+ # constant values at all. It is step 2 of the ratified sequence, written
2948+ # before any training corpus exists, so every claim it makes is checkable
2949+ # here on both platforms and none of it needs a measurement.
2950+ - name : Calibration policy controls (no defaults, symmetry, exact fit)
2951+ run : python tests/test_calibration_policy.py
2952+ # Step 4 froze WHICH BYTES the policy is, and nothing else. These controls
2953+ # recompute that digest from git blobs at the frozen commit, and refuse the
2954+ # artifact outright if it ever grows a field that is not identity.
2955+ - name : Calibration policy freeze controls (step 4 digest identity)
2956+ run : python tests/test_calibration_freeze.py
2957+ # Step 5 records the five constants the owner chose, bound to the step 4
2958+ # identity. The frozen implementation is the judge of its own ranges, so
2959+ # these controls hand it the committed pairs rather than re-stating them.
2960+ - name : Calibration design constants controls (step 5 ratified values)
2961+ run : python tests/test_calibration_constants.py
2962+ # Step 6 preregisters a training collection that has not happened. These
2963+ # controls prove the protocol is complete and bound, and that the scope
2964+ # combiner reads the frozen selector's own margin instead of copying it.
2965+ - name : Training preregistration controls (step 6 protocol + scope binding)
2966+ run : python tests/test_training_prereg.py
2967+ # Step 7 needs two dedicated measurement hosts that do not exist yet. This
2968+ # captures WHAT a machine is and never how fast it is: no clock, no resource
2969+ # observation, proved by walking the AST rather than the prose. It also
2970+ # proves the tool sits outside all three frozen source sets, so preparing
2971+ # step 7 cannot move a digest steps 4, 5 and 6 are already bound to.
2972+ - name : Step 7 environment capture controls (measurement-free identity manifest)
2973+ run : python tests/test_step7_envcapture.py
2974+ # The same tool run for real on this runner. It writes nothing, starts no
2975+ # clock, and is NOT a binding input: a hosted runner is not admissible for
2976+ # step 7, and the manifest records the runner and CI flag that say so.
2977+ - name : Step 7 environment capture selftest (writes nothing, starts nothing)
2978+ run : python scripts/step7/envcapture.py --selftest
2979+ # Plan mode builds the arms, preflights THOSE bytes, freezes their
2980+ # identities and fixes the execution order — then returns before the
2981+ # timing function is reachable. Linux only: the arms are ELF binaries.
2982+ - name : Round 7 runner — plan mode, no clock started
2983+ if : matrix.os == 'ubuntu-latest'
2984+ run : |
2985+ python scripts/round7/runner.py \
2986+ --candidate "$PWD/rust/target/release/own-cli" \
2987+ --plan --build-dir "$RUNNER_TEMP/round7-arms" \
2988+ --out "$RUNNER_TEMP/round7-plan.json"
2989+ # Decisive workloads: availability and pins only. No clock, no resource
2990+ # sampling — this step is what "enumerated, hashed, availability-checked"
2991+ # looks like when it is the only thing permitted.
2992+ - name : Decisive workloads — non-timed structural smoke
2993+ run : python scripts/perf_baseline.py --smoke-decisive
2994+ # The timed legacy calibration pair USED TO RUN HERE, and it is gone while
2995+ # step 7 is closed. It started a real stopwatch, produced forty timed cells
2996+ # per platform on every push and uploaded them, which made the step 6
2997+ # preregistration's own claim to take no new observation false on the very
2998+ # commit that made it. A preregistration that fixes "exactly one collection"
2999+ # while CI quietly gathers numbers every twelve minutes has built an
3000+ # epistemic side channel, whatever anyone intends to do with the output.
3001+ #
3002+ # What still proves the apparatus, all of it untimed: --selftest above, the
3003+ # instrument controls, the round 7 apparatus controls, the round 7 runner in
3004+ # plan mode, the decisive structural smoke, and the calibration policy,
3005+ # freeze, constants and training preregistration controls.
3006+ #
3007+ # training-no-incidental-measurement fails this suite if any workflow
3008+ # regains a path to a measurement entrypoint while the preregistration
3009+ # still records step 7 as unauthorised.
3010+
28973011 # P-014 Tier B: external-reference resolution. The SAME sample, run two ways, must give two
28983012 # verdicts — proving the extractor binds a THIRD-PARTY event only when its DLL is referenced:
28993013 # A (no refs) -> ObservableObject is an error type -> OWN050 (honest skip), no leak
0 commit comments