feat(step7): host qualification and the execution binding, kept apart

[PhysShell]

Phase 2 of the S8 repair chain. Stacked on #355 (the instrument's memory-metric repair), which it needs for the closed memory vocabulary.

Three artifacts, three questions, never merged:

	question
host qualification	does this environment satisfy the host predicate of a named, frozen T0?
session eligibility	is this session, on that qualified host, able to start now?
session admissibility	did the attempt that ran remain the one that was authorised?

and, separately, the execution binding: which Linux and Windows environments, which candidates and which instrument form this campaign. hostqual.py neither issues nor owns it — a utility that checks a CPU governor must not become the root of campaign identity. D7 later binds execution_binding_sha256 rather than restating the machines.

Qualification is versioned by T0

A qualification claims "this host satisfies T0-7", so it cannot float free of T0. --qualify requires the T0 path and commit and proves: the commit exists, blob_sha really is commit:path, the bytes hash to the recorded sha256, and the document declares FROZEN. A qualification against a NOT_FROZEN T0 is refused — today's reconnaissance cannot become campaign evidence by reuse — and the binding refuses a host qualified under a different T0 than it binds.

Every consumed artifact is proved before it is read

One boundary, one validator per artifact type: parse → prove kind/schema/shape → only then read semantics. Binding, qualification, environment manifest, preflight record, provisioning and session declaration all go through it.

Artifact validity is not predicate outcome. Malformed input — wrong kind, wrong schema, the string "false" where a boolean belongs, a VM omitting a VM-only field — is refused before any record exists. A declared boolean false is not malformed: it is a perfectly good declaration that this host does not qualify, and it reaches a real artifact. The exit codes say which happened:

0  valid artifact, positive outcome
1  valid artifact, NEGATIVE outcome — the record exists and names the failed check
2  malformed input or operational misuse — no record is produced

The envcapture schema string is duplicated rather than imported (neither tool imports the instrument or the capture tool), with a control proving the copy still equals what the producer emits.

Evidence classes stay apart

Declared — provisioning and the per-session operator declaration. Content-addressed, checked by shape and value, never called machine proof: a guest OS cannot establish that no neighbour arrived on the same hypervisor. dedicated_to_p022 and no_concurrent_user_workload must be true, hosted_ci_runner false, is_vm a real boolean; a VM must promise fixed vCPU, fixed RAM, no live migration and no dynamic memory, with n/a unavailable to it, while a physical host answers those four with an explicit n/a: <reason>.

Machine-observed — manifest.provenance.ci == false (the field always exists, so demanding its absence could never be satisfied); performance governor on every applicable CPU plus an identified turbo mechanism; on Windows an accepted plan with processor state pinned at 100% on AC and DC, read by GUID and by position because every label on a non-English machine is localized; and the stratum's memory mechanism actually existing here.

Identity, candidate, quiesce

• The qualification records a canonical hash of the whole identity block, and a session must reproduce it — a changed kernel, CPU count, RAM or toolchain can no longer walk past a matching fingerprint.
• The candidate is hashed against the binding before the clock, and again after the session.
• Quiesce is 120 s and is actually waited: 60 s quiet, then twelve 5 s intervals sampled from /proc/stat or GetSystemTimes, mean below 5 %, no interval above 20 %. A missing sample, a rewound counter or a zero denominator is NOT_ELIGIBLE rather than a skipped interval. A quiesce failure starts no clock, so it is not INVALID and costs no retry budget.
• A postflight pass re-checks identity, power and candidate and requires closing-probe evidence. Preflight may not certify what a session did after it started.
• A preflight and a postflight carrying two different execution_binding_sha256 values cannot meet in one session.

The binding proves its inputs, and --verify is total

The harness digest is recomputed from the instrument sources at the bound commit by the frozen formula, without importing the harness; because the workload manifest is one of those sources, proving the digest at a commit also proves the manifest there. --verify re-proves T0, instrument, manifest, both qualifications and both candidates — and there is no partial mode under that name: incomplete inputs are refused by the function and by the command line, so the string binding verified cannot appear over a skipped component.

Driven live, not only by fixture

The first live Windows run found a defect in this tool's own parser: it read the possible-range minimum as the current setting and reported 0% on a machine pinned at 100%. Fixed — the block ends with the two current indices, AC then DC.

Exploratory witness, not qualification evidence: on the development machine processor_min_dc = 5. Under the AC+DC rule that host does not qualify, which is the point of the rule — a machine can look perfect on its current power source and carry a different profile the moment the source changes.

Windows S606, honest declaration (this box is NOT dedicated):
  ci                     pass  manifest.provenance.ci == false
  power_policy           pass  active plan 8c5e7fda-…, processor state 100%
  required_memory_metric pass  max_process_peak_commit via win32 job object
  single_tenant          fail  the declaration is valid evidence that this host
                               does not qualify: dedicated_to_p022 is declared false
  => a qualification artifact IS written, qualified: false, exit 1
Linux (WSL):
  power_policy           fail  no cpufreq/scaling_governor on any CPU
  => artifact written, qualified: false

Both refusals are correct, and both leave evidence: neither environment is a measurement host, the tool says so for the right reason, and the attempt is recorded rather than discarded.

step 7 host qualification controls: 24 passed, 0 failed. Regression: envcapture 11/11, perf instrument 16/16, calibration freeze 7/7, training preregistration 9/9.

No real campaign qualification or binding has been created. T0 remains NOT_FROZEN; collection_authorized: false.

🤖 Generated with Claude Code

https://claude.ai/code/session_018xhcg5opoFbSTdYHpkSXCh

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Advanced

Run ID: a5634186-e63e-4640-a488-b824d382fab6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}