refactor(exaforce): isolate schema pruning as runtime monkeypatch

[will-exaforce]

Summary

• Reproduces this PR's original LLM schema/prompt pruning (fewer keys → smaller structured-output requests → fewer LLM timeouts) without editing any upstream-tracked source file, so future syncs from NVIDIA/SkillSpector merge cleanly.
• The four source files this PR originally edited are reverted to upstream parity; the pruning now happens at runtime from a fork-only src/skillspector/exaforce/ package applied on import skillspector.
• Every mutation is guarded — if an upstream rename/rewrite removes a target field or prompt sentence, import skillspector fails loudly with PatchDriftError instead of silently going stale.
• Net runtime behavior is identical to the original direct-edit version of this PR (same pruned schema, same trimmed prompts), so the attached benchmark comparison still applies.

Details

Why a monkeypatch instead of direct edits. Editing upstream files (llm_analyzer_base.py, meta_analyzer.py, the two semantic analyzers) produces a merge conflict on every upstream sync. Moving the change into the fork-only src/skillspector/exaforce/ package keeps those files byte-identical to upstream — the net diff vs main touches no upstream source/test file except a 4-line activation block in src/skillspector/__init__.py.

Mechanism. exaforce.apply_patches() mutates the Pydantic models in place — del model.model_fields[...] then model_rebuild(force=True) on both the leaf model and its container (the container caches the nested schema) — and trims three prompt module-globals via a guarded str.replace(). Mutating in place rather than swapping classes preserves class identity, so response_schema bindings and isinstance checks keep working untouched.

• LLMFinding (_schema_patches.py) — drops explanation, remediation, and stops forwarding them in to_finding
• MetaAnalyzerFinding — drops intent, impact
• MetaAnalyzerResult — drops overall_assessment and its _parse_stringified_assessment validator
• Both semantic ANALYZER_PROMPTs + the meta PER_FILE_ANALYSIS_PROMPT (_prompt_patches.py) — trim the now-unused line-number / intent / impact instructions

Activation. A two-line block at the end of src/skillspector/__init__.py calls apply_patches() after from skillspector.graph import ... (so every target module is already loaded — no circular import). Pruning is therefore the default on every entry path: CLI, MCP server, and the benchmark's library import.

Why two upstream test files fail — by design. Because the upstream test files are also kept at parity, tests/nodes/test_llm_analyzer_base.py and tests/nodes/test_semantic_quality_policy.py assert the un-pruned schema and so fail at runtime (4 tests). This is a deliberate trade: full upstream parity was chosen over a green upstream suite, to avoid the ongoing burden of keeping forked test assertions in sync with upstream. The exact expected-failure set is recorded in docs/superpowers/EXPECTED_TEST_FAILURES.md as a regression tripwire — a failure there is only a real problem if it is not one of those four pruned-key assertions (e.g. an import/collection error). The fork-only tests/exaforce/ suite covers the patch behavior itself.

Benchmark. Runtime behavior is unchanged from the original direct-edit version of this PR, so the earlier main-vs-PR comparison still holds: report.md

Tests: fork-only tests/exaforce/ suite 10/10; full suite 4 failed (all documented in EXPECTED_TEST_FAILURES.md) / 1261 passed.

Implementation Plan

Reproduce this PR's runtime effect from an isolated exaforce package while restoring every upstream file to parity.

#	Layer	File	Change
1	Revert	4 source + 2 test files	Restore to pre-PR upstream parity
2	Patch lib	src/skillspector/exaforce/_patchlib.py	Guarded primitives (raise PatchDriftError on drift)
3	Schema patch	src/skillspector/exaforce/_schema_patches.py	Prune model fields + model_rebuild
4	Prompt patch	src/skillspector/exaforce/_prompt_patches.py	Trim prompt module-globals
5	Orchestrator	src/skillspector/exaforce/__init__.py	Idempotent apply_patches() (_PATCHED flag)
6	Activation	src/skillspector/__init__.py	2-line apply_patches() call on import
7	Tests	tests/exaforce/**	Guard-primitive unit tests + subprocess patch/activation tests
8	Docs	docs/superpowers/EXPECTED_TEST_FAILURES.md	Documented expected-failure set

Verification: git diff <upstream> -- <the 6 files> is empty (parity); the tests/exaforce/ suite is green; the 4 full-suite failures match EXPECTED_TEST_FAILURES.md and are bounded to the two reverted files.

[smoy]

you may still want to follow up to figure out how we should have a sanity CI test suite, just so we know we didn't accidentally break something.