PEP 705 ReadOnly TypedDict inheritance (E0038): conformance 136→137, false positives 126→120, mutation-hardened + de-flaked nvim CI gate

[MelbourneDeveloper]

TLDR

Implements PEP 705 ReadOnly TypedDict inheritance — flipping typeddicts_readonly_inheritance.py to PASS (conformance 136→137 = 93.84%) while dropping total false positives 126→120; also adds clean profiler/debugger demo scripts under examples/ and de-flakes the Neovim e2e CI gate.

What Was Added?

• crates/basilisk-resolver/src/scope/typeddict_meta.rs — shared, cross-crate TypedDict membership primitives: is_transitive_typeddict, has_extra_items_transitive, transitive_typeddict_names, strip_typeddict_qualifiers, class_by_name. These exist because ClassInfo::is_typed_dict is only true for classes that name TypedDict directly, so a subclass (class Album(NamedDict): ...) was invisible to every TypedDict rule — the single root cause of both the missed diagnostics and the E0014 false positives.
• crates/basilisk-resolver/src/visitor/typeddict_schema.rs — effective_fields, which merges a TypedDict's own + inherited fields (most-derived declaration wins) carrying each field's ReadOnly qualifier.
• E0038 redeclaration-legality matrix (crates/basilisk-checker/src/rules/e0038.rs) — new pure decision functions: parse_field_qualifiers, redeclaration_violation, value_type_incompatible, type_head, is_invariant_container, bases_conflict.
• 5 benchmark stress fixtures (auto-discovered by benchmarks/run.sh): e0038_typeddict_readonly_inheritance.py (the new rule), plus e0056, e0093, e0050, e0036 — each ~2000 lines and verified to fire its rule 840–4000×.
• 8 mutation-safe tests for E0038 in crates/basilisk-checker/tests/mutation_kill_tests.rs.
• Profiler/debugger example scripts (examples/cpu_demo.py, examples/memory_demo.py, examples/README.md) — clean, fully-typed scripts meant to be run (F5) rather than statically checked: cpu_demo.py produces a lopsided CPU flame chart (hot is_prime / deep fib_recursive / asleep cold_io); memory_demo.py exercises a sustained leak, a transient spike, and a __del__ reference cycle for the memory profiler; the README catalogues every example with its expected diagnostics.
• assert_plenary_pass (scripts/common.sh) — a result-based gate for the Neovim e2e suite (see "Changed" below), plus spec section [LSPTEST-EDITOR-SPECIFIC-INTEGRATION-NEOVIM-E2E-GATE] in docs/specs/LSP-TEST-INTEGRATION-SPEC.md.
• Spec section [CHKARCH-DIAG-TYPEDDICT-READONLY-INHERITANCE] in docs/specs/CHECKER-ARCHITECTURE-SPEC.md.

What Was Changed or Deleted?

• E0038 no longer skips qualifier-wrapped fields. It enforces PEP 705: a writable item may not be redeclared ReadOnly; a required item may not be redeclared not-required; writable value types are invariant while ReadOnly value types may narrow to a subtype (same invariant container — list/dict/set — with different args is rejected, a different container head is allowed); multiple inheritance with conflicting core type / required-ness / read-only-ness is rejected. Field-override diagnostics now point at the field, not the class name.
• E0014 (e0014/mod.rs) recognises transitive TypedDict subclasses so it skips their dict-literal assignments (field-level checking belongs to E0093) — this clears the FP cluster across the read-only suite.
• E0056 / E0093 (resolver final_readonly.rs, typeddict.rs, typeddict_ext.rs, core.rs) operate on the effective merged schema: inherited-ReadOnly writes are flagged, and wrong-type / missing-required checks run against inherited fields including plain reassignment. del td[k] and the mutating dict methods now honour extra_items (PEP 728).
• Dedup: removed strip_td_wrappers/try_strip_wrapper (resolver) and is_typed_dict_class (checker), now sharing strip_typeddict_qualifiers / is_transitive_typeddict.
• Neovim e2e CI gate (scripts/test-nvim.sh): the PlenaryBustedDirectory run is now gated on parsed results rather than the nvim process exit code. Under make ci's parallel -j3 load the PlenaryBustedDirectory parent can exit non-zero on teardown (a late-reaped LSP child / async handle) even when every test passed — a flaky false failure. assert_plenary_pass instead requires that every spec file started AND emitted a summary with zero failures, zero errors, and no Lua traceback — strictly stronger than the exit code (it also catches a run that silently executed no tests, which the old gate passed).
• Ratchet: coverage-thresholds.json FP ceiling 126→120 (decreases monotonically); conformance_status.csv regenerated (137/146).

How Do The Automated Tests Prove It Works?

• Conformance gate (conformance_tests::conformance_score): typeddicts_readonly_inheritance.py now PASS (caught 11, missed 0, FP 0); scorecard reports 137/146 and False+: 120; both gates print Conformance gate: 93% … PASS and FP gate: 120 <= 120 ceiling — PASS.
• Mutation testing (cargo-mutants, scoped to the 8 E0038 functions): 33/33 viable mutants killed (1 non-compiling/unviable). The 8 #[mutation_safe(rule = "e0038", …)] tests (mutant_e0038_parse_readonly, …_required_relaxing, …_writable_invariant, …_invariant_container, …_type_head, …_field_override, …_bases_conflict, …_single_base_no_conflict) each assert both a violating redeclaration that MUST fire and a legal one that MUST NOT — including a 3-base case that kills the len() < 2 boundary mutant.
• Coverage thresholds enforced and met: basilisk-checker 94% ≥ 93%, basilisk-resolver 96% ≥ 94%, all crates green (scripts/test-rust.sh).
• Existing e0038_tests, e0056, e0093 integration suites and all 125 checker_tests pass unchanged.
• Full make ci is green end-to-end (exit 0): lint + Rust tests/coverage + VSIX (328 passing, 87% ≥ 86%) + Neovim e2e (16/16 files ran, 16 summaries, 0 failed, 0 errored, coverage 44% ≥ 30%) + build. The Neovim gate itself was verified against five inputs — it passes a clean run yet fails on an injected test failure, an errored test, a missing-summary (truncated) run, and a Lua traceback.

Spec / Doc Changes

• docs/specs/CHECKER-ARCHITECTURE-SPEC.md: new #### ReadOnly TypedDict inheritance {#CHKARCH-DIAG-TYPEDDICT-READONLY-INHERITANCE} section; E0038/resolver code and the mutation tests reference this spec ID.
• docs/plans/CHECKER-PEP-CONFORMANCE-PLAN.md: marks typeddicts_readonly_inheritance.py DONE; score 137/146.
• CLAUDE.md: codifies the monotonic conformance / FP / benchmark policy (pre-existing staged guideline edit carried in this branch).

Breaking Changes

• None — additive diagnostics on previously-unchecked TypedDict redeclarations and new examples/ scripts; no public API removed (resolver gained exports only).