Never serve a masked decode from a staged speculation fire

[shsym]

Problem

Pass-level speculation pre-fires each context's next decode step before the inferlet has set up that step's constraints, so every staged fire is built with an empty constraint surface: no logit mask (build_next_request emits logit_masks = []) and no custom attention mask (masks = [], has_user_mask = false). The staged-hit check entry_matches_request compared anchor token/position and the spec arrays but neither mask — so a request carrying one was served the unconstrained, fully-causal staged token verbatim, silently dropping the constraint at the unconstrained→constrained transition.

Two reachable variants, both default-on (speculation_depth = 1):

• Logit mask — json_object / json_schema / forced tool-call grammar decode ignored the grammar. A real model hides this (asked for JSON it naturally emits a leading {); the dummy driver (uniform-random within the mask) exposes it as fully unconstrained output that never begins with {.
• Attention mask — single-token windowed-attention / attention-sink / hierarchical-attention decode (these decode one token at a time and call attention_mask() once the sequence exceeds the window) was served a token computed with full causal attention, so it attended to positions the inferlet meant to exclude.

Fix

Reject a staged hit whenever the incoming request carries either mask, forcing it onto the cold path where the mask is applied (the try_hit miss branch clears the now-stale chain). This mirrors the producer-side skip set in evaluate_request_shape, which already refuses to stage from either kind of masked request — the two sites must stay symmetric or the drop reappears.

A regression test per variant proves a masked request misses an otherwise-matching staged entry (both mutation-proven: each fails if its mask clause is reverted).

Also teaches the dummy driver to ack Copy/Adapter as no-ops (status 0) instead of the catch-all -1 error: a no-compute driver has no KV pages to copy and no adapter weights to load, and token generation is independent of cache contents, so a vacuous success is faithful. The previous -1 broke the context-fork path used for prefix sharing.

Provenance

Integration bug surfaced by the reconcile: this branch's runtime/src/inference/speculator.rs (pass-level speculation) and the pie-bridge logit_masks/masks schema were developed on separate upstream branches and were not wired together until the reconcile merged them onto one tree. Neither source branch exercised both halves, so the staged-fire-vs-masked-request interaction was never tested until now.

Verification

Tier	Result
inference::speculator::tests (10)	green — incl. entry_match_rejects_logit_masked_request + entry_match_rejects_user_masked_request, both mutation-proven
HTTP chat E2E	202 passed, 1 skipped (#200 pre-existing); json_object/json_schema output begins with {
Real Metal	json_schema → 200 with {"name": ...}; forced tool → 200 with tool_calls; S3 coherence green
pie lib / weight-loader	444 / 14

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: a51705a0-5835-43f4-a582-eaae29e51915

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch scratch/655-spec-mask-fix

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}