Runs, jobs, and sweeps

The authorization chain

prepare_strategy_run freezes the exact draft, artifact, validation, dataset, runtime (version-file hash + git commit), profile, timeout, seed, analyzers, and parameter grid into a hash-bound plan and returns a signed run token. A separate local approval (backtrader-mcp approve --run-plan ... --yes) is mandatory before start_strategy_run, which consumes the approval and the token nonce after the durable job row exists.

Run profiles

Profile	Modes	Purpose
runonce	runonce	Fast single pass
runnext	runnext	Incremental pass
runonce_runnext_compare	runonce, runnext	Determinism gate
fixed_tests	runonce, runnext	Default; determinism gate
parameter_sweep	runonce per combination	Typed parameter grid

parameter_sweep freezes a param_grid (StrategySpec parameter names to value lists, at most 64 combinations) under the same single approval. Parameters are passed through cerebro.addstrategy(..., **override); results carry a sweep block with every combination's metrics ranked by return_rate.

Job lifecycle

States: QUEUED, RUNNING, CANCEL_REQUESTED, CANCELLED, SUCCEEDED, FAILED, TIMED_OUT, ORPHANED. Every transition is a compare-and-swap write with one arbitration rule: a terminal state, once persisted, is never overwritten, and a visible CANCEL_REQUESTED suppresses SUCCEEDED/FAILED/TIMED_OUT.

A server-owned watchdog (started only by serve) consumes the worker heartbeat, enforces the wall-clock deadline with a grace period, orphans jobs whose worker died, and cleans up detached candidate process groups. Jobs report a structured error_kind (user_strategy/resource_limit/timeout/validation/infrastructure/ cancelled/orphaned).

The concurrency cap rejects instead of queueing: start_strategy_run fails with an actionable suggestion at max_concurrent_jobs.

Results and comparison

Successful results contain exactly eleven canonical metrics (bar_num, buy_count, sell_count, win_count, loss_count, trade_num, final_value, sharpe_ratio, annual_return, max_drawdown, return_rate); sharpe_ratio and annual_return are nullable. The packaged comparison-profile-v1 policy is the single authority for comparison tolerances, including the tightened final_value override (rel 1e-9 / abs 1e-6). Allowlisted analyzer metrics appear under extra_metrics; every run manifest fingerprints the runtime commit and the resolved pandas/numpy versions.