Add session-scoped context experiments

[2001Y]

Summary

• add a session-scoped context_experiments config block for A/B testing prompt context arms before system-prompt assembly
• support deterministic hash or persisted round-robin assignment by session id, with optional platform filters
• load experiment arm context files and optional preloaded skills in place of normal project context
• document the config shape in prompt assembly docs

Tests

• scripts/run_tests.sh tests/agent/test_context_experiments.py tests/agent/test_system_prompt.py tests/agent/test_prompt_builder.py tests/hermes_cli/test_prompt_size.py tests/hermes_cli/test_prompt_compose_command.py -q
• .venv/bin/python -m ruff check agent/context_experiments.py agent/prompt_builder.py agent/system_prompt.py tests/agent/test_context_experiments.py tests/agent/test_system_prompt.py
• .venv/bin/ty check agent/context_experiments.py tests/agent/test_context_experiments.py tests/agent/test_system_prompt.py

Notes

• A full scripts/run_tests.sh attempt initially failed because the newly created local .venv lacked pytest; after syncing the dev extra, the targeted canonical suite above passed.

[teknium1]

Thanks for the session-scoped assignment design and the focused tests. I found one cross-platform correctness issue.

Problems

• agent/context_experiments.py:67-75 — the msvcrt branch opens an empty a+ lock file and, on OSError, logs then executes the state read-modify-write cycle without a lock. This breaks the stated round-robin serialization guarantee for concurrent Windows processes. Existing lock implementations document and handle this requirement by seeding one byte, opening r+, and seeking to offset zero before msvcrt.locking() (hermes_cli/auth.py:1024-1037; tools/skill_usage.py:99-108).

Suggested changes

• Seed and seek the Windows lock file using the established helper pattern, and do not perform the assignment update when locking cannot be acquired.
• Add coverage for the msvcrt path and lock failure/concurrent assignment behavior; the new tests currently exercise only serial direct builder calls (tests/agent/test_context_experiments.py:43-78).

Automated hermes-sweeper review.

[teknium1]

On Windows, msvcrt.locking() requires a non-empty lock file and a file pointer at offset zero; this a+ open leaves the new lock file empty, and the exception path below then continues the round-robin read-modify-write cycle unlocked. Please follow the seeded-byte + r+ + seek(0) pattern in hermes_cli/auth.py:1024-1037, and avoid assigning when the process lock cannot be acquired.

[2001Y]

Rechecked the current head e5eb2a6 against the context-experiments review feedback.

The current implementation and regression coverage are present; no additional code change was needed in this recheck.

Verification:

• python -m pytest -q tests/agent/test_context_experiments.py
• git diff --check

Please re-review the current head.