perf: incremental parent-map maintenance

[itamarga]

TLDR

Fixes a quadratic parent-map rebuild that made pyjsclear.deobfuscate() run for 15+ minutes on large obfuscator.io-style samples. On two representative samples, 0.1.5 takes 1011.9s / 1134.9s; this branch takes 10.4s / 27.5s under the same profiler (3.1s / 8.3s without).

Changes

• base.py: new record_replacement(replacement, parent, key, index) — an O(1) patch of the cached parent map instead of a full invalidation.
• ObjectSimplifier, ClassStaticResolver, StringRevealer: node replacement now calls record_replacement(...) instead of invalidate_parent_map(), eliminating the full O(N) parent-map rebuild on every subsequent find_parent.
• Pin esprima2>=…,<5.0.2: 5.0.2 switched optional-chaining (a?.b) parsing to the ESTree ChainExpression wrapper, which the generator and transforms don't yet support (full support tracked separately).
• Version bumped to 0.1.6.
• New/updated tests for the incremental parent map.

Full description

Root cause. Every node replacement in ObjectSimplifier, ClassStaticResolver, and StringRevealer called invalidate_parent_map(), so the next find_parent rebuilt the entire parent map with a full-AST walk. With thousands of proxy-object references that is O(n²). Profiling 0.1.5 on the 606KB sample shows build_parent_map called 12,023 times, consuming 980.8s of the 1011.9s total (97%) — 1.76 billion dict.get calls — all triggered by node replacement invalidating the parent map on every swap.

Evidence (same machine, both versions under cProfile):

Sample	pyjsclear 0.1.5	this branch (cProfile)	this branch (no profiler)
606KB obfuscator.io-style JS	1011.9s	10.4s	3.1s
753KB obfuscator.io-style JS	1134.9s	27.5s	8.3s

Fix. record_replacement(replacement, parent, key, index) patches the cached parent map in O(1). This is safe because the replacement swaps a node in place: in-place swaps don't shift list indices, and replaced subtrees are detached and never looked up again, so only the replacement's own entry changes. The incremental map is behaviourally identical to a full rebuild.

esprima2 pin. esprima2 5.0.2 changed optional-chaining parsing from a bare MemberExpression (with optional: true) to the ESTree-spec ChainExpression wrapper. The code generator has no ChainExpression handler and the transforms don't see through the wrapper, so a fresh install that floats the dependency up to 5.0.2+ breaks optional-chaining round-trips and leaves obfuscated identifiers unresolved. Capping below 5.0.2 restores a green build; full ChainExpression support (generator + transforms) is tracked as a separate follow-up.

Tests. Full suite green: 1754 passed, 3 skipped (unit + regression, fuzz excluded). The parent-map test asserts build_parent_map is called exactly once across many in-place replacements — a deterministic check of the fix rather than a timing threshold.

🤖 Generated with Claude Code

[itamarga]

Superseded by #14 — reopened from a renamed branch (perf/incremental-parent-map). Same commits, no content change.