Make inspect and language-server timeouts configurable

[mathiasloh]

Problem

When running aft_inspect on a large monorepo, the language server may need more than two minutes to start and finish publishing diagnostics.

At present, blocking diagnostics have a fixed 120-second deadline. If the language server is still loading when that deadline expires, AFT returns:

lsp_quiescence_timeout

The user cannot increase this limit through configuration.

There is also a separate transport timeout for the inspect command. These two timeouts can be misaligned: the transport may stop waiting before the internal diagnostics phase has finished. Increasing only the transport timeout exposes the internal 120-second failure but does not resolve it.

Current implementation

The relevant values are currently set here:

• The blocking diagnostics phase has a hard-coded 120-second deadline in diagnostics_category.rs (BLOCKING_DIAGNOSTICS_PHASE_TIMEOUT).
• The bridge transport timeout for inspect is set to 60 seconds in command-timeouts.ts (inspect: 60_000).

Proposed change

Make the blocking diagnostics timeout configurable, for example:

{
  "inspect": {
    "blocking_diagnostics_timeout_ms": 300000
  }
}

The current default could remain 120000 so existing installations keep their current behavior.

The inspect transport timeout should also be configurable or automatically derived from the diagnostics timeout, with enough additional time for the final response. For example:

{
  "bridge": {
    "inspect_request_timeout_ms": 330000
  }
}

Alternatively, a single inspect timeout setting could control both values and prevent the transport timeout from being shorter than the internal diagnostics deadline.

Expected behavior

• Small repositories continue to use the current defaults.
• Large repositories can increase the timeout without rebuilding AFT.
• AFT still reports a clear timeout failure when the configured limit is reached.
• Invalid or excessively small timeout values are rejected with a useful configuration error.
• The configured timeout is covered by tests for both the diagnostics phase and the transport layer.

A configurable timeout would allow users to accommodate slow language-server startup while preserving the existing bounded behavior.

[alfonso-aft]

Accepted, both halves. You've correctly identified that there are two misaligned budgets: the hard-coded 120s BLOCKING_DIAGNOSTICS_PHASE_TIMEOUT and the transport-side request budget that gives up first (we have this banked internally as a known residual from the v0.51 blocking-fresh work — the server often finishes computing a terminal exactly as the client abandons it).

The fix lands as one unit:

1. A config key for the blocking diagnostics phase deadline (user tier, project-overridable upward) so large monorepos can raise it past 120s.
2. Wait-aware transport budget for aft_inspect: the plugin maps a transport timeout derived from the configured phase deadline (the same pattern our bash wait:true path uses), so the client never abandons a request the server is still legitimately serving.

One heads-up while you're on a large monorepo: #242/#243 track the fact that a long blocking inspect currently occupies its execution lane while it waits, which can queue parallel sessions' mutations behind it — raising the timeout will make that window longer until that redesign ships. Both are moving now.

[alfonso-aft]

Both halves are on main (d7a16857), landing in the next patch release:

1. inspect.diagnostics_timeout_ms — user tier, default 120000, clamped 10000..600000. A project config may raise the effective value but never lower it (a project must not silently shrink another consumer's diagnostic completeness). Documented in docs/config.md.
2. Wait-aware transport budget — both plugins now derive the aft_inspect transport timeout from the configured deadline plus 30s headroom, so the client outlives the server's legitimate work instead of abandoning at a fixed 60s while the server finishes computing the answer.

For your large-monorepo case: set the key at user tier (~/.config/cortexkit/aft.jsonc) to e.g. {"inspect": {"diagnostics_timeout_ms": 300000}} and the transport follows automatically.

The related lane-occupancy issue (#242) is in design review — until that ships, a longer deadline does mean a longer window where parallel sessions' mutations queue behind the inspect, so on shared roots it's worth raising deliberately rather than to the ceiling.

[alfonso-aft]

Shipped in v0.51.3.