feat(config): require an explicit context window instead of assuming 16384

[Max17190]

Why

context_tokens defaulted to 16384, a window sized for the small local models the loop was first hardened against. Everything derived from it (the transcript budget, the compaction trigger, the max_tokens clamp) then treated a 128k or 500k model as a 16k one, so the harness's own context protection became the constraint on the models it now targets. A larger default is wrong the other way for a small model, and nothing is queried from the server, so the honest answer is the one base_url and model already give: no default.

Summary

• Settings.context_tokens is Option<usize> with no default; old files still parse.
• providers::resolve returns ResolveError::MissingContextTokens when neither the model's providers.json entry nor settings.json sets a window; the message names both places and the launch-read rule. Every surface that shows resolve errors (turn start, -p/--stdio fail-fast, /status, /provider) inherits it.
• openmax --check now warns on a settings file that parses but cannot resolve (endpoint, model, window, provider), and reports the resolved window on a healthy one.
• README, docs/configuration.md, config.rs module doc, and --spec settings updated; the README example sets context_tokens.

Test Plan

• New a_context_window_is_required_from_settings_or_the_model_entry: flat and named-provider resolution refuse without a window, a per-model entry satisfies it, settings.json satisfies it for an unsized model.
• check_names_a_settings_file_that_would_brick_the_next_launch extended: warns with the reason when the window is missing, prints the window when resolvable.
• Every fixture that relied on the old default now sets a window explicitly (unit, integration, and CLI tests).
• End to end on the built binary: -p refuses with the message; --check prints the warn line.
• cargo test --workspace green; clippy at zero warnings.

Note for reviewers: users upgrading with a settings.json that omits context_tokens get a one-time resolve error naming the field; the external rigs already set it.

Greptile Summary

This change removes the assumed context-window size and requires each model to declare its actual window through settings or a provider model entry. A local endpoint check confirmed that a model-specific provider window starts successfully, while unsized configurations stop before sending a model request.

Confidence Score: 5/5

No blocking failure remains.

No accepted blocking findings remain. The exercised configuration paths correctly allow a provider model's declared window and reject missing windows before network use.

T-Rex Logs

What T-Rex did

• Built the open-max-tui package with cargo --locked and executed the context-window harness against the built binary.
• Observed that the harness produced a streamed response and emitted 'no context window configured' for missing-window scenarios, with the probe staying on a single request.
• Ran the unit test for the open-max-core context-window requirement; the test passed.
• Reviewed the CLI harness source and the two context-window logs to verify the observed behavior.

_{Ran code and verified through T-Rex}

_{Reviews (2): Last reviewed commit: "feat(config): require an explicit contex..." | Re-trigger Greptile}