fix(tools): read preprocessor branch combinations lazily in import_audit - #11622

@MarcZX10 Thanks, you're right: each group's alternatives were still built in full before the first combination came out. 3d2aebfa03 keeps nested groups lazy as well. Each group's alternatives are now produced only when the walk reaches them, in the same order as before, and the limit counts only complete combinations.

The test now includes your layout: an outer #ifdef whose first branch holds the nine #ifdef blocks, with the stub call in #else. It is reported as partial, and the same test fails on 94d991e03a with more than 256 preprocessor branch combinations. A nested function that reaches a stub in every combination still fails. I also compared the new branch_variants with main's on 60,000 random nested bodies: every one gives the same combinations in the same order, or fails the limit in both. The description has the updated results.

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

A real-world check of tools/import_audit.py from this PR's head 3d2aebfa03, run against current main 62b3ba2522 sources.

Setup: converted import registries (*.registry.json from relinker --windows --registry) of 22 titles, read from local copies of their dumps. --libs is a Windows Release build of main 3fea2a937b. Windows 11, Python 3.15.1. Only tools/import_audit.py was swapped in; the other sources stay at main.

• main's import_audit.py: all 22 runs stop with FAIL: more than 256 preprocessor branch combinations in one function, before auditing anything.
• This PR's import_audit.py: all 22 runs complete and print the per-class counts, with no FAIL line. For example, PPSA19591 gives 1074 implemented, 29 partial, 5 stub and 6 absent imports.

I did not compare classifications line by line with an older main. This only shows that the tool runs again on real registries.

AI-assisted: yes (Claude Code).

Independent Linux confirmation of your 3d2aebfa0313a518a43530f2a9da777f196cb330, tested in a private merge with main 2ee8e9751a06bd4cb7edb8d17f433d56a39a7025 (bd494620dc498c979601c8048e46f726d5e02308).

Ubuntu 26.04.1 x86-64, Python 3.14.4:

• Full import-audit suite: 155 tests run, 154 pass, one Windows-only skip. Conventions and whitespace pass.
• Unchanged main's scan of its actual core/libs/prx stops at poll_nid_postfix in libkernel/Socket/src/Udp.cpp with the 256-combination refusal. This candidate completes the scan. All 7,225 stub entries and 524 partial entries match exactly against the isolated reference using unchanged main code with a process-local limit of 4,096. The limit change is only a validation control, not the production fix. The complete normalized result SHA256 is e05bc24d738f3ac5a49f0278b67b54d8f1b25b5e4c2fa61d17e59cca19137df4.
• Replayed both saved Middle Ages Ping Pong PPSA28319 datasets (October 17/23), using the actual pinned-current-main classification sources, all five registries, three module directories and hash-checked frozen runtime libraries. Main exits 2 before reporting; this candidate completes each audit and exits 1 with the same separate four-import module-identity problem tracked by fix(tools): recognize guest module identities in import audit #6171. That remaining exit 1 is expected and does not mean the source-scan correction failed.
• A separate private integration with the reviewed identity/path corrections and fix(tools): accept PT_DYNAMIC data after DT_NULL in import_audit #11499/fix(tools): require a module suffix on both names when matching aliases #11581 reaches exit 0 on those saved datasets and retains the same complete source classifications. Its full tools suite runs 1,367 tests: 1,342 pass, 25 environment/platform skips. These combined results do not describe the standalone fix(tools): read preprocessor branch combinations lazily in import_audit #11622 tree.

These are fresh Linux Python/static-audit checks using archived converted assets, not new title, native-game, Windows or console execution. The runtime libraries retain their original October 17/23 build pins; no PR branch has been changed.

AI-assisted validation and reporting: OpenAI Codex.

@BunnyFl Independent Windows confirmation of 3d2aebfa0313a518a43530f2a9da777f196cb330, cleanly merged with current main f759e165d9cc3088b196f9b544e038687003fef4 in validation composition 3a1e9245cd43. Windows 11 Enterprise 26200, Python 3.13.15.

• Import-audit, progress, progress-wrapper and NID-name suites: 256 tests run, 254 pass, two POSIX-permission skips. Conventions and whitespace pass.
• Keeping your regression while substituting main's production module fails with the expected more than 256 preprocessor branch combinations error. The candidate passes, including its flat/nested always-stub cases that retain the refusal.
• The candidate completes classification of the actual current-main core/libs/prx: 7,232 stub-or-partial entries, 528 partial. It also completes our separately pinned Sonic integration 879d6d9ba22f45932a9cebc35bdc682d9fb7b388: 7,231 entries, 527 partial. Complete normalized classification hashes and timings are in the receipt.

Reproduction script, test logs, negative control and exact pins. These are public-source and synthetic-fixture checks, not fresh game execution, a full import-registry audit, or a claim that classifications were compared with an older main. No branch limit was raised. I encountered the same poll_nid_postfix failure independently; your existing PR covers it, so no competing PR is being submitted.

AI-assisted validation and reporting: OpenAI Codex.

Checked the lazy expansion (nested alternatives and the complete-combination limit), that variant_calls consumers (extend, short-circuiting all) read it once, and test_import_audit (155 passed, 1 skipped); merged.