feat(benchmarks): add concurrent-write driver

[phernandez]

Summary

Ports the concurrent-write convergence benchmark from basic-memory-benchmarks into Core's canonical /benchmarks package.

• Adds deterministic multi-session MCP create/edit workloads and marker-based convergence checks.
• Records per-operation latency/errors plus file, entity, observation, relation, and lost/doubled-write integrity.
• Shares the warm MCP runtime with the existing local retrieval provider.
• Updates the benchmark docs and command surface for the Core-owned package.

Why

The benchmark suite now lives in Core, while standalone PR basic-memory-benchmarks#42 was opened against the old repository. Keeping the driver there would strand issue #1248 as that repository is deprecated.

This port also fixes all three unresolved review findings from the standalone PR:

1. The convergence verdict is captured before the optional full reindex mutates derived state.
2. A timed-out MCP request is cancelled, its stdio subprocess is terminated by the transport, and the client thread must exit before verification continues.
3. Concurrent-write runs require a local git checkout and record its exact resolved SHA.

Implementation

• Adds bm_runtime.py for an explicitly managed warm MCP stdio session.
• Adds concurrent_write.py for deterministic planning, concurrent execution, settle detection, integrity checks, and artifacts.
• Adds bm-bench run concurrent-write plus smoke/load just recipes.
• Refactors bm-local to use the shared MCP runtime.
• Documents Core as the canonical benchmark location and requires BM_LOCAL_PATH for reproducible concurrent-write runs.
• Rejects dirty target checkouts and calculates throughput from concurrent operations only, excluding sequential hub setup.

Verification

• just test — 168 tests passed.
• just lint — passed.
• Changed-file pyright — 0 errors.
• just --list — passed; both concurrent-write recipes parse.
• Real MCP smoke on the final fix: 4 writers, 100 notes, 4 hubs — converged with 194/194 operations and every integrity check passing before reindex.
• Package-wide pyright still reports two pre-existing errors in test_baseline_providers.py and test_mem0_normalization.py; neither file is changed here.

Risk

The main behavioral change outside the new command is the shared lifecycle wrapper used by bm-local. Its normal startup/call/stop contract is unchanged; shutdown is stricter because it now fails instead of silently leaving a live session behind.

Refs #1248

[phernandez]

@codex review

[phernandez]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 37b61771d2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[phernandez]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 365bf175bb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[phernandez]

@codex review

[phernandez]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b60830c250

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Resolve the benchmark SHA from the harness checkout

When bm-bench is invoked from outside the benchmark checkout, git_sha(Path(".")) records the caller's repository HEAD—or "unknown" outside a Git repository—rather than the SHA of the benchmark code being executed. This silently misattributes published comparison artifacts; resolve the repository containing the harness package and fail if its exact revision cannot be captured.

AGENTS.md reference: benchmarks/AGENTS.md:L82-L85

Useful? React with 👍 / 👎.