EXPERIMENT: elide Python attribute after CFG nodes

[yoff]

Experimental DCA only

This draft isolates one Python shared-CFG representation experiment: merge ordinary AttributeExpr after-completion with the canonical attribute operation while retaining the operation and its exception/abrupt-completion origin.

Not proposed for merge. The purpose is to measure fleet heterogeneity and memory effects from removing a high-frequency positional CFG role.

Exact topology

• Baseline: f7774efbc4502d015f21e001cd399d61f33cc9d9
• Candidate: 1fd2a6c5b971f70c1c3baa259a9dd80d6c037a78
• Sealed patch SHA-256: a54b1ffcded04ae408bab4e025fceede18479908a875e15d5124d2877d6b99f3
• One experimental commit; two production files; no unrelated diagnostics.

Local controls

NiceGUI:

• 21,645 fewer CFG/positional nodes;
• five fewer blocks;
• complete lifecycle: 91,437,534 -> 91,166,650 joined tuples, -270,884 (-0.296%);
• rung-A shared-CFG construction: +39,003 joined tuples;
• normal-clean cache: 31,404 -> 30,464 KiB;
• tuple pool unchanged at 4,952 KiB.

Additional symmetric controls:

• Langroid: -19,469 nodes, -0.193% lifecycle joins, rung A +51,079, cache -1,268 KiB;
• DeepDiff: -4,044 nodes, -0.059% lifecycle joins, rung A +28,990, cache -376 KiB;
• no tuple-pool growth.

Correctness gate

• focused shared CFG/SSA tests: 8/8 passed;
• focused Python dataflow tests: 14/14 passed;
• exact complete canonical adjacency on NiceGUI, Langroid, and DeepDiff;
• exact complete nonempty public local-flow relations on all three corpora;
• SSA definitions, phis, first-use, and adjacent-use unchanged;
• production MissingCallToDel BQRS byte-identical on all three corpora.

The known first-construction regression is caused by the generic AttributeExpr membership route; later stages save work because the reachable graph is smaller. The DCA tests whether corpus variation or space effects make this tradeoff useful anywhere beyond the pinned local controls.

[yoff]

DCA launched: github/codeql-dca-main#39772

• Baseline: f7774efbc4502d015f21e001cd399d61f33cc9d9
• Candidate: 1fd2a6c5b971f70c1c3baa259a9dd80d6c037a78
• Toolchain / extractor / compiler / evaluator: be44d8072494633e7caf0a79bae276a4c5090c35
• DCA revision: 00dae12ec8cdb10aab3fc9065a109841acc13860
• Sources: exact canonical 48-source Python set from github/codeql-dca-main#38848
• Suite/source configuration: nightly.qls / nightly.yml
• Database cache: disabled
• Grouping/action/toolchain/source inputs: structurally identical to the reference after normalizing experiment/variant/bundle identities
• Controller: Tasks run 37544982718
• Reports: data/yoff/PR-160-python-attribute-after/reports

The local token is not SAML-authorized for the worker organization, so dispatch used DCA’s supported --skip-assert-worker-repository-access preflight exception. The generated worker inputs were unchanged. No follow-up DCA will be launched while this experiment is unresolved.

[yoff]

DCA completed: github/codeql-dca-main#39772

• Coverage: 54/54 successful analyze runs for each variant; 0 failed targets
• Correctness: 0 changed code-scanning rows, 0 added/removed alerts, 0 changed query-result tuple counts
• Total query execution time: 6,404s -> 6,425s, +0.328%
• Complete-source analysis sum: +0.326%; paired-source median: 0.000%
• Sources: 21 improved, 7 unchanged, 20 regressed
• Cache after normal cleanup: -0.778%
• Tuple-pool cache: -0.858%
• Database after normal cleanup: -0.156%
• DIL: +0.0661%

Conclusion: REJECT the exact experiment as a performance change. The retained-space reduction is real, but fleet runtime is neutral/slightly worse, deterministic local gains are only 0.059-0.296%, rung-A construction regresses, and generic hook machinery expands DIL. No follow-up DCA will be launched without explicit authorization.