# T2.3 — Plumb BWE into `AdaptiveQualityController` **Status:** Pending Review **Agent:** Kimi Code CLI **Started:** 2026-05-11T17:13Z **Completed:** 2026-05-11T17:20Z **Commit:** 846c98e **PRD:** ../PRD-transport-feedback-bwe.md ## What I changed - `crates/wzp-proto/src/quality.rs` — `AdaptiveQualityController`: - Added `bwe: Option>` field. - Added `set_bandwidth_estimator(&mut self, bwe: Arc)` setter. - Added `tier_ceiling_bps(tier: Tier) -> u64` helper using `QualityProfile::total_bitrate_kbps()`. - In `try_transition()`, before upgrading to a higher tier, check BWE headroom: ```rust if let Some(ref bwe) = self.bwe { let required = (Self::tier_ceiling_bps(next_tier) * 130) / 100; if bwe.target_send_bps() < required { self.consecutive_up = 0; return None; } } ``` This requires `target_send_bps() >= 130%` of the next tier's bitrate ceiling (including FEC overhead). ## Why these choices The 130% headroom factor is a safety margin: we only upgrade if the bandwidth estimate comfortably exceeds the target tier's requirement, preventing flapping when BWE is borderline. Resetting `consecutive_up` to 0 on BWE block gives the estimator time to converge before the next upgrade attempt. Checking the *next* tier's ceiling (not the current tier) is correct: the guard answers "can we afford the tier we're trying to move into?" ## Deviations from the task spec None. ## Verification output ```bash $ cargo test -p wzp-proto quality running 24 tests ...(all 24 pass)... test result: ok. 24 passed; 0 failed; 0 ignored; 0 measured; 95 filtered out; finished in 0.10s ``` ## Test summary - Tests added: 1 - `bwe_guard_blocks_upgrade_when_bandwidth_insufficient` — verifies low BWE blocks upgrade, high BWE allows it after counter reset - Tests modified: 0 - `wzp-proto` test count: 119 (was 118 before T2.3) - `cargo clippy -p wzp-proto --all-targets -- -D warnings`: pass - `cargo fmt --all -- --check`: pass ## Risks / follow-ups - `BandwidthEstimator` is attached via `set_bandwidth_estimator()`; call sites in `wzp-client` (send/recv loops) will create and wire it in a future task. - The BWE guard only applies to upgrades. Downgrades are unchanged — they react quickly to quality reports regardless of BWE. ## Reviewer checklist (filled in by reviewer) - [ ] Code matches PRD intent - [ ] Verification output is real - [ ] No backward-incompat surprises - [ ] Tests cover the new behavior - [ ] Approved