perf(cypher): VarMap replaces Binding HashMap for clone-per-match path (FU-006 H2)

[coseto6125]

Summary

Third and final fix from the FU-006 follow-up spike's top-3 hotspot list. PR #432 attacked the iteration count (CSR kind lookup + walk_rel closure); this PR attacks the per-iteration clone cost. They compose orthogonally.

The cypher executor's Binding row holds three name→index maps that are cloned once per matched node in exec_pattern's frontier expansion. The old HashMap<String, u32> clone was the dominant cost — 67-85% of query time per the spike — because each clone heap-allocates a fresh bucket table and per-key String.

Replace node_vars and edge_vars with VarMap: a SmallVec of (CompactString, u32) entries that stays inline for the typical 2-4 vars per cypher query. Inline storage means Binding.clone() is a fixed-size memcpy with zero heap allocations on the hot path. CompactString keys inline up to 24 bytes (every realistic cypher var name), so the per-key clone is also alloc-free. Linear scan over ≤4 entries beats HashMap lookup at this size because there is no hashing cost.

Wider Binding API kept compatible — get(key: &str) -> Option<&u32>, contains_key(key: &str) -> bool, insert(key: &str, value: u32) — identical to the HashMap surface so existing call sites just drop the now-redundant var.clone() before insert. Five sites updated.

The computed field stays HashMap<String, Value>. It carries non-Copy Value items and is populated only by WITH clauses, never during frontier expansion. Empty-HashMap clone is a 56-byte memcpy with no allocation, so the savings would not justify the Option<Box<HashMap>> ergonomics churn.

Empirical (--runs 10, freshly-rebuilt .sample_repo, same-session A/B vs main)

Phase	Main	H2	Δ
cypher count(*) ungrouped	61.4 ms	35.6 ms	-42%
cypher decorator IN	48.8 ms	41.4 ms	-15%
find (bm25)	22.1 ms	18.9 ms	-14%
inspect (Class)	23.3 ms	20.3 ms	-13%
impact upstream	18.3 ms	16.0 ms	-13%
impact downstream	18.1 ms	16.0 ms	-12%
summary	19.9 ms	17.8 ms	-11%
summary --detailed	19.6 ms	17.4 ms	-11%
routes	12.7 ms	11.4 ms	-10%
cypher Method-Calls→Method	74.4 ms	73.3 ms	-1%
cypher Class→Method	45.0 ms	47.9 ms	+6% (noise)

count(*) is the headline (~26 ms saved on the 110k-Method frontier-expansion clone path); the consistent -10% to -14% across other query shapes comes from the same Binding clone being on more code paths than just exec_pattern (inspect / impact / find all build bindings during their internal traversals).

Composes with PR #432 (kind-CSR + walk_rel closure)

Query	Main	#432 H1+H3	This H2	Stacked expected
cypher count(*)	61.4 ms	32.4 ms (-47%)	35.6 ms (-42%)	~20 ms (-67%)
cypher decorator IN	48.8 ms	39.1 ms (-20%)	41.4 ms (-15%)	~33 ms (-33%)

H1+H3 cuts iteration count from 303k → 110k; H2 cuts per-iteration clone from 74 ns → ~5 ns. Both halve their respective axis, so combined should compound. Will validate empirically once #432 lands and this rebases on top.

Spike attribution

Spike report identified all three hotspots; this PR completes the trio:

• H1 (kind-CSR scan): shipped in PR perf(cypher): exec_pattern kind-CSR scan + walk_rel closure (FU-006 follow-up) #432
• H2 (Binding SmallVec): this PR
• H3 (walk_rel closure): shipped in PR perf(cypher): exec_pattern kind-CSR scan + walk_rel closure (FU-006 follow-up) #432

Test plan

• cargo test -p egent-code-plexus --tests — 1182 passed
• cargo test -p ecp-core --tests — 268 passed
• cargo clippy -p egent-code-plexus --tests -- -D warnings — clean
• cargo clippy -p ecp-core --tests -- -D warnings — clean
• Same-session A/B against main on 11-phase benchmark — zero regression, -10% to -42% wins
• Manual cross-check that the apparent impact --baseline HEAD~1 slowdown in the bench was a bench-state confound (both main and H2 take ~1.4s when L1 needs refresh; the bench's 9 ms reading on main was from a previously-warm L1 state)

[github-actions]

ecp impact cache (0 symbols) — internal, used by ecp dev pr-analyze

[]