[codex] add headless training command

README.md

@@ -102,6 +102,7 @@ bun run demo:drone
 
 The `init:*` commands create template projects without running experiments. The `demo:*` commands create timestamped projects under `.ignitionrl/demos/` and write export artifacts under each project `exports/` directory.
 Existing projects can be inspected without rerunning experiments through `bun run --cwd packages/cli start inspect <projectDir> --json`.
+Registered environments can be trained without the demo shortcut through `bun run --cwd packages/cli start train <projectDir> <environmentId> --learner tabular-q --episodes 30 --checkpoint-id final --json`, which persists traces, metrics and a learner checkpoint.
 Persisted runs can be compared without rerunning experiments through `bun run --cwd packages/cli start compare <projectDir> --score-by summary.totalReward --json`.
 Metric series can be inspected through `bun run --cwd packages/cli start metrics <projectDir> totalReward --json`.
 The offline Studio bootstrap also embeds metric chart summaries for persisted metrics, so the future metrics panel can render chart selectors and latest values immediately.

packages/cli/README.md

@@ -24,6 +24,12 @@ bun run --cwd packages/cli start demo drone-target ./drone-target-demo.ignitionr
   --learner-episodes 12 \
   --inference-episodes 3
 
+bun run --cwd packages/cli start train ./grid-world-project.ignitionrl GridWorld-v0 \
+  --learner tabular-q \
+  --episodes 30 \
+  --checkpoint-id final \
+  --json
+
 bun run --cwd packages/cli start inspect ./target-2d-demo.ignitionrl \
   --score-by summary.totalReward \
   --json
@@ -124,6 +130,8 @@ If the project directory is omitted, the CLI creates a timestamped project under
 
 Each demo command runs the environment, writes traces, metrics and learner checkpoints when available, then exports JSON artifacts under the project `exports/` directory. The DroneTarget demo also replays the final continuous-control checkpoint as a separate inference run so the learned policy is directly comparable and replayable.
 
+`train` opens an existing local project, verifies that the selected environment is registered, runs one learner, and writes traces, metrics and a checkpoint. Supported first-class learner selections are `tabular-q` and `dqn` for `GridWorld-v0`/`Target2D-v0`, and `linear-policy-search` for `DroneTarget-v0`. Use `--json` for automation and `--checkpoint-id` to choose the checkpoint name.
+
 - project report;
 - Studio project view;
 - Studio environment view when `environment --export` is used;