fix(tasks): deadlock is a catchable runtime error at main's join site (#509)

[InauguralPhysicist]

A cooperative-task deadlock (all tasks blocked, none runnable) was raised inside scheduler_trampoline via rt_error and surfaced as a fatal process error — a try/catch around task_join never ran the handler. Yet deadlock is a documented closed error kind, and a killed task's interrupt on join is catchable. This closes that inconsistency.

Fix

The trampoline detects deadlock only when the main task is still live and suspended (the DONE/DEAD case returns earlier). At that point it now builds the structured deadlock error at main's blocked line and:

• If any of main's saved frames is inside a try → clear main's block reason and re-enqueue main; the scheduler loop resumes it with g_has_error set, so CHECK_ERROR unwinds to the handler. e.kind == "deadlock", e.message == "all tasks are blocked — deadlock", and execution continues after the block.
• Otherwise → terminal, exactly as before: loud message + non-zero exit.

The deadlock is delivered to the main task — a try inside a worker does not catch it.

define da as: return task_join of gb
define db as: return task_join of ga
ga is task_spawn of da
gb is task_spawn of db
try:
    task_join of ga        # before: fatal, handler skipped
catch e:
    print of e.kind        # after: "deadlock"
print of "after"           # after: runs; rc 0

Implementation note (a latent bug this surfaced)

The fatal print no longer routes through rt_error's g_try_depth gate. g_try_depth is a process global, not part of a task's saved slice, so a suspended worker's still-open try can leave it non-zero between task switches — which silently suppressed the fatal deadlock print when a worker held a try. The trampoline now decides catchable-vs-terminal from main's saved frames directly and prints the terminal case itself.

Tests

• Catchable-deadlock assertion at the end of test_tasks.eigs — red without the fix (the uncaught deadlock aborts the whole file before it can print its summary). Asserts e.kind == "deadlock", the message, and that code after the try runs.
• task_deadlock_worker_try.eigs in the exit-contract harness — a worker's try does not catch (rc 1, stays fatal).
• Existing task_deadlock.eigs gate still asserts uncaught deadlock is loud + rc 1 and leak-clean.

Docs

SPEC.md (task layer), BUILTINS.md (task_join), CHANGELOG [Unreleased].

Validation

Release + ASan suites 2688/2688, leak tally 0, 0 heap errors. Every new assertion confirmed red-without-fix.

Closes #509

🤖 Generated with Claude Code