Add CodSpeed performance benchmarks

[codspeed-hq]

Summary

This PR sets up continuous performance measurement with CodSpeed for the tokscale-core crate, which holds the CPU-bound hot paths of the project (session parsing and model aggregation).

Changes

• Added codspeed-criterion-compat as a dev-dependency of tokscale-core (imported under the criterion name), so existing criterion-style benchmarks integrate directly with CodSpeed.
• Added crates/tokscale-core/benches/core_benchmarks.rs covering hot paths exercised on every scanned message:
  • normalize_model_for_grouping over a representative batch of model identifiers (Anthropic/GPT canonicalization, custom prefixes, date suffixes, reasoning tiers, free markers).
  • SIMD-accelerated JSONL parsing (parse_jsonl_file) at 100 / 1000 / 5000 lines to capture scaling behavior.
  • JSON array parsing (parse_json_file) over 1000 objects.
• Added a dedicated [profile.bench] to the workspace Cargo.toml. Benchmark builds inherit from release, which enables full LTO and a single codegen unit; on this dependency tree (bundled SQLite, vendored TLS, zstd) that made benchmark compilation prohibitively slow. The bench profile keeps opt-level = 3 while disabling cross-crate LTO and allowing parallel codegen units. Release binaries are unaffected.
• Added .github/workflows/codspeed.yml running benchmarks in CPU simulation mode on pushes to main, pull requests, and manual dispatch, authenticated via OpenID Connect (id-token: write).
• Added the CodSpeed badge to README.md.

Notes on runners

The workflow uses CPU simulation mode on ubuntu-latest, which is deterministic and hardware-agnostic, so standard GitHub-hosted runners are appropriate here.

Verification

Benchmarks were built and run locally in simulation mode (cargo codspeed build + codspeed run --mode simulation -- cargo codspeed run); all five benchmarks were measured and results uploaded successfully.

Next steps

• Merge to establish the baseline on main.
• Consider adding benchmarks for the aggregation/sessionize pipelines once representative fixtures are wired up.

---

Summary by cubic

Adds CodSpeed performance benchmarks and CI to track tokscale-core hot paths (model normalization and JSON/JSONL parsing). Bench builds are faster with a custom profile; release builds are unchanged.

• New Features

  • Added CodSpeed-compatible benches for normalize_model_for_grouping, parse_jsonl_file (100/1000/5000 lines), and parse_json_file (1000 objects).
  • Added GitHub Action to build and run cargo-codspeed on push, PR, and manual trigger in deterministic simulation mode.
  • Added [profile.bench] to speed bench compilation (keeps opt-level = 3, disables LTO, sets codegen-units = 16).
  • Added CodSpeed badge to the README.

• Dependencies

  • Added dev-dependency codspeed-criterion-compat (aliased as criterion) and a bench target for tokscale-core.

^{Written for commit 1bec0ee. Summary will update on new commits.}

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 5 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• mixed_batch: 13.5 µs
• parse_json_file/1000_objects: 1.4 ms
• parse_jsonl_file[100]: 218.1 µs
• parse_jsonl_file[1000]: 2.2 ms
• parse_jsonl_file[5000]: 10.9 ms

---

[makoMakoGo]

Closing this CodSpeed wizard PR because PR #70 has already integrated the useful benchmark/profile/badge pieces while preserving the fork-specific workflow hardening and personal/local-clients triggers. The generated workflow here should not be merged.