[belikh]

The async REX3 shader compiler takes work from one bounded channel (256 slots). A full channel drops the request so the GFIFO consumer never blocks, but it cannot tell a warm-up/prefetch shape from the one the guest is drawing right now. A boot-time profile flood can fill every slot and starve the live draw, which then runs on the interpreter until the backlog drains — 30-second waits in graphics tests.

• Split the work into two lanes. The cold lane (warm-up, prefetch) keeps the 256-request bound and stays best-effort: a rejected request drops its Queued marker so a later draw retries.
• The hot lane is the draw path. It is unbounded, never rejected, and the worker drains it before the cold lane, so live draw work cannot be starved by prefetch.
• The worker parks on a condvar rather than spinning. The queue carries no Cranelift types, so the lane behaviour (including a hot request jumping a full cold queue) is unit-tested without starting a compiler.
• clanker/rex3/rex-jit-queue-retry.md records the drop-and-retry contract.

Tests: cargo test --workspace on upstream/main — 1332 passed, 0 failed.

[belikh]

Evidence behind the change (the fork now requires this for every upstream PR):

• Reproduction. A 263-entry warm-up profile fills the old single 256-slot channel. The next live draw calls request_compile; try_send fails, the Queued marker is removed, and that draw runs on the interpreter until the backlog drains. On the fork this showed up as 30-second waits in graphics tests with a populated ~/.iris/rex-jit-profile.bin. With the two lanes, the same profile's live draw is admitted to the hot lane and the cold backlog drains behind it.
• Tests in this PR. hot_draw_request_jumps_a_full_cold_queue fills the cold lane and requires a hot draw request to be admitted and drained first; hot_request_is_admitted_ahead_of_a_full_cold_queue checks the admission order at the queue level without starting a compiler.
• Contract. clanker/rex3/rex-jit-queue-retry.md (updated here) records the drop-and-retry rule for a rejected cold request.

If you would prefer the 263-entry profile as an in-tree regression fixture rather than a documented reproduction, say so and I'll add it.

[techomancer]

ps after reconsideration this is probably a no-op nowadays. 500 or so rex3 shaders that are most commonly used are already monomorphised draw functions and are not using the compile queue. there are probably very little compile requests at all. we may want to revisit it and clean it up.