test: add benchmarks, examples, and bench-regression guard (#15)

[millerjp]

Summary

Completes the Phase 4 benchmark scaffolding: a full per-method benchmark suite, overhead pairs vs raw sync.Map, runnable godoc Examples for every public symbol, the committed bench.txt baseline, and the benchstat-regression-guard CI job (mirroring mask).

Closes #15.

Highlights

• syncmap_bench_test.go — expanded from 9 to 28 benchmarks. Covers Load (hit + miss), Store, LoadOrStore (loaded + stored), LoadAndDelete, Delete, Range, Len, Map, Keys, Values, a concurrent 90/10 read-write pattern (pre-computed key pool so allocs/op isn't dominated by the harness), and five Overhead pairs (Load/Store/LoadOrStore/Delete/LoadAndDelete each with syncmap vs raw sync.Map).
• example_test.go — 14 runnable Example* functions, one per public symbol. Iterators (Range/Keys/Values/Map) sort output so // Output: blocks are deterministic. All pass go test -run Example.
• bench.txt — committed baseline from darwin/arm64, -count=5. Header documents regen procedure and the CI consumer.
• scripts/check-bench-regression.sh — ported from mask with two fixes: skip geomean summary rows and exclude zero-delta matches. The script is now stable against itself (a bench.txt → bench.txt compare reports zero regressions, which mask's original would false-positive on).
• Makefile — bench-regression target.
• .github/workflows/ci.yml — benchstat-regression-guard job (benchstat pinned to same version as mask, Dependabot skip, shared-runner CAVEAT preserved). makefile-targets-guard expected list updated to include bench-regression.
• .gitignore — bench-regression.txt, current.txt.

Local validation

$ make check                      # 14/14 gates green, coverage 100%
$ make bench-regression           # "No regressions above threshold."
$ go test -run Example ./...      # 14/14 pass
$ actionlint .github/workflows/*.yml  # clean

Agent gates

• Plan — single-commit scope, six-file bundle, bench.txt committed from this machine (matches mask's M2 baseline convention).
• performance-reviewer (post) — no BLOCKING. Applied both IMPORTANTs: concurrent bench pre-computes keys (dropping a spurious 1 alloc/op); Overhead pair comment clarifies the Store-included workload.
• code-reviewer (post) — no BLOCKING. Six NITs, all optional or already addressed.
• go-quality — 8/8 PASS.
• commit-message-reviewer — PASS.

Deferrals

• CHANGELOG — docs: add README, CHANGELOG, NOTICE, SECURITY, and LLM docs bundle #17.
• README "Performance" section linking to bench.txt — docs: add README, CHANGELOG, NOTICE, SECURITY, and LLM docs bundle #17.
• CONTRIBUTING §Performance baseline (referenced by the CI job comment) — docs: add CONTRIBUTING, CLA, CODE_OF_CONDUCT, CONTRIBUTORS and enable CLA/contributors workflows #18 (noted there).

Test plan

• All 13 CI jobs green, including the new benchstat-regression-guard
• benchstat-report artefact uploaded from the guard job
• Coverage remains 100%