--timeout semantics silently changed from "compile-only" to "whole command", breaking existing runs without any error output (3.0.63 → 3.0.64)

[descampsk]

Summary

Between 3.0.63 and 3.0.64, the --timeout flag on dataform run was redefined without a major version bump:

• 3.0.63: "Duration to allow project compilation to complete." — only bounded the compile phase.
• 3.0.64: "Wall-clock deadline for the entire command. When it fires, in-flight work (including any running JiT compilation worker) is cancelled." — now bounds compile and the full execution of every action in the run.

This change appears to have shipped as part of #2211 ("Integrate JiT compilation into CLI runtime"), which introduced the new per-model --jit-timeout flag alongside a silent redefinition of the pre-existing --timeout flag.

Any project that was passing --timeout to dataform run under the old (compile-only) semantics — a very reasonable thing to do, since a short timeout is sensible for compilation but not for execution — now has that same value applied as a hard deadline for the whole run. Once it fires, in-flight actions are cancelled and all not-yet-started actions are marked "skipped", with no error message printed at all: no exception, no stack trace, just a silent early exit (exit code 1) partway through the action graph.

Impact

• Runs that used to complete now stop partway through with zero explanation, unless you already know to compare the last completed action against elapsed wall-clock time.
• This can affect production runs equally — a scheduled dataform run --timeout 30s (fine when --timeout was compile-only) will now truncate any real run that takes longer than 30s end-to-end, silently dropping tables/assertions that were never
  reached.
• No deprecation warning, no changelog entry we could find calling out this behavior change, and no major version bump despite it being a breaking change in flag semantics.

Steps to reproduce

1. On 3.0.63, run a project with dataform run --timeout 30s where full execution (compile + all BigQuery jobs) takes longer than 30 seconds. Run completes successfully.
2. Upgrade to 3.0.64 only (no other changes). Run the exact same command.
3. Once 30s of wall-clock time elapses, the run stops: the in-flight action is dropped, all remaining pending actions are logged as Skipping dataset creation / Skipping assertion execution, and the process exits with code 1 — with no error, warning, or timeout message printed anywhere in the output.

Expected behavior

• This should have been called out explicitly as a breaking change (changelog + version bump), since it changes what a previously-passed flag value does to already-shipped automation.
• When the wall-clock deadline fires and cancels in-flight/pending work, the CLI should print a clear message (e.g. Run cancelled: --timeout of 30s exceeded) rather than exiting silently with skipped actions and no explanation. Right now nothing in run.ts's cancellation path surfaces this to the console — it looks like the parent CLI simply doesn't log anything when CANCEL_EVENT/onCancel fires the way it does when an action genuinely fails.

Version

• Introduced in: 3.0.64
• Last known good: 3.0.63
• Suspected source: Integrate JiT compilation into CLI runtime #2211