Fix diagnostic pressure oracle norm range

[MrScripty]

The diagnostic pressure oracle rejects finite, structurally verified schema-2 pressure comparisons around 1e-200 and 1e200: squaring unscaled samples underflows or overflows even when the L2 norm is representable. Scale each norm by the largest absolute sample before the independent math.fsum reduction. Metric definitions, source/selection/gauge checks and the existing 2e-13 relative / 1e-300 absolute tolerances remain unchanged. Nonfinite inputs and unrepresentable norms fail explicitly.

Stacked on #95 at exact 34f3d1dd0599bab199c003aae8a2c7f457eb70b0. Only the Python tool and one new automatically discovered desktop Cargo integration contract change. No existing Rust implementation, workflow, source inventory, admission record, provider or accepted archive semantics change. Lead retains integration and all merges.

Validation on frozen files now committed at d58044a9496a6abe39e568fd301d2de31e693dec:

• A real public DiagnosticRecorder closes each manufactured finite-pressure fixture; cold DiagnosticReader.verify and exact selection succeed. Rust metrics match the independent integer sum-of-squares closed form for 24 cells at tiny, large, ordinary and zero scales.
• The new Cargo contract passes 12 positive and 60 independently perturbed metric checks across normal Python, -O and PYTHONOPTIMIZE=1. Each negative must reach its named metric refusal. Existing desktop Cargo CI discovers the new test without workflow changes.
• The same test source compiled against the exact baseline oracle fails at the intended tiny-case L2 check. No source swapping or test bypass hook was used.
• Independent GPT-6.1 Sol review reran the Cargo contract and approved exact oracle SHA256 afa9845b1513a4119777f9564e5020e8503e0eabaed897aead8e2a11e9f5a299 and test b492194a384779fd15ba967bac93a5068d1635d5ff5b443d13a31252716134d7. Focused locked Clippy with warnings denied, formatting and whitespace checks pass.
• The actual native-inspector exported 4,096-cell teacher fixture summary also passes the fixed oracle in all three Python modes.

Evidence: /tmp/rheon-diagnostic-pressure-oracle-freeze.json, /tmp/rheon-oracle-independent-review.log, /tmp/rheon-oracle-baseline/artifacts.json. Fresh isolated /tmp/rheon-oracle-fix-target excludes earlier cross-worktree shared-target contamination. Test pressure/host observations are manufactured storage facts; the consumed manifest wrapper is deliberately minimal and is not full FrozenRunSet admission evidence. This proves reduction/custody behavior, not model accuracy, physical acceptance, runtime/hardware qualification or full M1–M3 acceptance.

Keep draft pending applicable exact-head CI and lead serial disposition. PR95's independently reported shared Rust custody, chronology and diagnostic-view findings have separate reviewable patch proposals; this tool fix does not close those findings. Preserve active dependent branches.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

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.}

[MrScripty]

All currently applicable hosted96 checks now pass at frozen head d58044a. Desktop job116681290554 checked actual test-merge919804f5a0ad9a2a3f6d255f1c3907b8a09fe96e (d58044a onto34f3d1d):725passed/0failed/5explicit ignores across overlapping suites. Its log explicitly runs closed_schema2_extremes_match_analytic_norms_and_oracle_in_every_python_mode and passes1/1 (25.25s), exercising the72 positive/perturbed calls. Existing measured-deployment ignores remain unqualified. Independent source review and local baseline-red/fixed-green proof remain as described. No merge performed; lead owns serial disposition with97/98.

[MrScripty]

Repository cleanup: closing this superseded scoped proposal. Its reviewed change is incorporated on main d8771ad83c30644d29929c96534548ce238bb357 through 8f30a004c1431fe11859da771179057fdaa0bfaf. Scaled finite L2 norm with independent fsum reduction and explicit unrepresentable rejection is retained. Both the Python oracle and the Cargo oracle-range test are byte-identical in main. Independent source/history review confirmed that the allocated product change is retained. The original branch is not classified as fully merged and is retained by this cleanup.