feat(hash): full bound/free identifier split, v2 recipe (#77)

[Connorrmcd6]

Summary

Closes #77. Completes the v2 hash recipe into the bound/free split — the principled fix for the gate's highest-value blind spot.

v1 alpha-renamed every identifier, so a one-token change re-pointing a span at a different single-occurrence external symbol (PointsTier.TIER_1 → TIER_2, getHighest → getLowest) produced an identical hash. check stayed green while the claim's prose silently became false.

What changed

• Bound/free split (surf-core/src/hash.rs). v2 alpha-renames only bound names — the symbol's own name, parameters, locals, loop/range/comprehension vars, with/catch aliases, generic params, destructuring binders — and emits every free identifier verbatim. Re-pointing at a different member, call target, type, enum/const, object key, or decorator is now loud even when it occurs once; consistent local renames stay quiet. Two-pass: collect_bound/bind_here (per-family binding tables) then emit.
• Subsumes two special cases. The check: member-access names verbatim in the hash (v2 recipe) — high-value slice of #77 #140 member-access-only first cut and the Python decorator rule (Python: decorator span gap when anchoring a function #8) both fall out — they're just free identifiers now. (One dedicated check remains so a member name stays verbatim when it collides with a bound local: x the param vs obj.x.)
• Recipe decision: completed v2, did not add v3. 0.7.0 is unreleased, so v2 is redefined in place. v1 stays byte-frozen — golden fixtures confirm released v1 stamps still verify.
• Fail-closed. tree-sitter only, no scope analysis: a position not positively a binding defaults to free. The one accepted approximation (match-arm pattern identifiers left free) is documented and pinned.
• Differential validation harness (surf-core/tests/differential_hash.rs, new). In-tree v1-vs-v2 A/B across all four languages, kept in-tree so any future canonicalization change reruns the same gate.
• Governance + docs. Re-pinned v2 golden digests; version table + N-1 support policy + identification/remedy in docs/reference/hash-recipes.md; updated how-it-works.md and CHANGELOG.md; dogfooded — claims in hubs/hash.md anchored to collect_bound/is_member_access_name, all hubs re-stamped.

The external git-history replay over real corpora (prometheus / nansen-python-sdk / a TS repo) named in the issue runs out-of-tree against release binaries; the in-tree harness is its deterministic, always-on counterpart.

Verification

All gates run on this branch:

• cargo fmt --all --check → clean
• cargo clippy --all-targets --all-features -- -D warnings → 0 warnings
• cargo test --all → all pass (incl. golden + differential)
• Differential harness: 13 benign / 0 regressions, 12 semantic / 100% v2 catch / 0% v1
• Dogfood surf check → all anchored spans match

🤖 Generated with Claude Code