@@ -1,6 +1,6 @@
 ---
 name: codetrial-rust
-description: How Rust is written and reviewed in CodeTrial - which rules the build enforces (the [lints] table in Cargo.toml and scripts/check-rust-style.py) and which it leaves to judgment, how to suppress a lint, why a lint is switched off on purpose, the file-size ceilings, and the ownership, error, cast and test idioms the tree follows. Use when writing, refactoring or reviewing Rust under src/, tests/, examples/ or build.rs, when clippy or the rust-style gate fails, when tempted to write #[allow], a bare unwrap, an `as` cast, a glob import or a bool parameter, when a file nears its line ceiling, or when asked to audit or tighten the Rust style. Comments, commit messages and what stays out of a public issue are codetrial-conventions; running the gate is codetrial-verify.
+description: How Rust is written and reviewed in CodeTrial - which rules the build enforces (the [lints] table in Cargo.toml and scripts/check-rust-style.py) and which it leaves to judgment, how to suppress a lint, why a lint is switched off on purpose, the file-size ceilings, and the ownership, error, cast and test idioms the tree follows. Use when writing, refactoring or reviewing Rust under src/, tests/ or build.rs, when clippy or the rust-style gate fails, when tempted to write #[allow], a bare unwrap, an `as` cast, a glob import or a bool parameter, when a file nears its line ceiling, or when asked to audit or tighten the Rust style. Comments, commit messages and what stays out of a public issue are codetrial-conventions; running the gate is codetrial-verify.
 ---
 
 # Writing Rust in CodeTrial
 
@@ -138,12 +138,6 @@ tokio = { version = "1.54.1", features = ["test-util"] }
 # line; the SDK already depends on this crate.
 log = "0.4.34"
 
-# The audio probe matches onsets to pulses, and `cargo test` builds an
-# example's tests without running them unless asked to.
-[[example]]
-name = "audio-latency"
-test = true
-
 # scripts/hook-check holds every `cargo build` to the git hooks being in
 # effect; its build.rs says why it is a member of its own. `vendor/` holds path
 # dependencies, not members, and `desktop/` is a crate of its own with its own
 
@@ -431,7 +431,6 @@ web/            Static browser application
 problem-bank/   Problem and judge source of truth
 config/         Env file templates and per-provider credentials
 docs/           Contracts and operational notes
-examples/       Standalone measurement probes built against the crate
 scripts/        Build, verification, and integration checks
 vendor/         Patched third-party crates
 tests/          Rust and browser tests
 
@@ -785,110 +785,3 @@ export file was named differently. CodeTrial keeps the tool split and the
 residual, joins asynchronous records by identity rather than by neighboring
 lines, and refuses an empty analysis. It uses milestones measured on its own
 clock instead of estimating provider stages it cannot see.
-
-## Synthetic LiveKit audio suite
-
-`python3 scripts/measure-audio-latency.py --output /tmp/audio-local` builds the
-`audio-latency` example against this checkout's locked dependencies and repeats
-three trials of the 100 ms and 20 ms native source queues, alternating their
-order. Each trial contains one labelled warmup pulse and twenty measured
-pulses. `--queues`, `--iterations`, and `--pulses` select the experiment. Output
-must be a new directory, created only after the build succeeds; an earlier
-experiment is never overwritten.
-
-The default `--mode local` connects a native source to a local track sink and
-needs no service. `--mode room` publishes to a fresh, uniquely named LiveKit
-room and measures the remote decoded PCM sink. Load `LIVEKIT_URL`,
-`LIVEKIT_API_KEY`, and `LIVEKIT_API_SECRET` into the environment first, using
-the selected private config. No Gemini requests, candidate media, or interview
-dispatch are involved. The room mode creates ordinary synthetic participants
-and consumes LiveKit connection minutes. On Linux x86_64 the runner uses
-`scripts/fetch-clang.sh` unless `CXX` already selects a compiler.
-
-The probe sends 24 kHz mono PCM in ten-millisecond frames, delivered in 100 ms
-bursts. A 100 ms, 1 kHz tone follows 300 ms of silence in each one-second trial
-window, and a final silent second follows the last pulse. RMS hysteresis detects
-an onset after decode. Both participants share one process and one monotonic
-clock, so host clock synchronization is not an assumption.
-
-An onset counts for a pulse only within the one-second delivery window after
-its send, which ends no later than the next pulse's send. An onset in a gap
-between the window's deadline and the next send is recorded as late. Without an
-embedded pulse identity, a delivery slower than the whole period can land in
-the next pulse's window and look fast there; the pulse it belongs to is then
-missing, and under a constant overrun the warmup is always one of them.
-Variable delays and losses can still alias individual samples; this suite does
-not prove pulse identity or measure latency beyond its matching window. Zero,
-multiple or late onsets make a sample unsuccessful, the warmup included. Raw
-records include the send offset, first tone-frame capture completion, detected
-delivery latency, and on-time and late onset counts. The probe writes them
-before leaving the room, so a capture timeout or a failed unpublish or close
-cannot discard what was measured; the file retains measurement errors and
-teardown errors separately, and any of them exits nonzero. A pending teardown
-marker is saved first, then updated after every unpublish and close has been
-attempted. Each cleanup operation has a five-second deadline; a failed save
-also cannot skip cleanup. The raw result uses schema version 3; earlier files
-without teardown status are rejected. This is an acoustic onset estimate with
-ten-millisecond frame resolution; codec ringing, loss concealment, scheduling
-and jitter buffering can affect detection. It is not sample-accurate waveform
-correlation.
-
-`local` measures source buffering and sink delivery; `room` additionally
-includes encoding, the network and SFU, and decode/jitter buffering. Neither
-measures a browser's speaker output, microphone hardware, Gemini inference,
-or production worker backlog. Capture completion is reported separately: it
-can precede delivery. In the pinned native implementation the configured
-source queue is a backpressure threshold; capacity is twice that threshold.
-Changing it can move waiting into the application worker instead of making
-the listener hear a reply sooner. Repeated trials are essential: an arbitrary
-phase relative to the native ten-millisecond timer alone can change onset
-latency by nearly a frame.
-
-Each experiment saves `manifest.json` with revision, whether tracked files
-differ from it, one digest over the manifest, lockfile, build script and every
-file under `src/` and `examples/`, the executable hash, the Rust and C++
-compiler versions, and host identity. Each trial saves raw JSON.
-`summary.text`, `summary.json`, and `summary.csv` pool the measured samples by
-mode and queue size, retaining attempted, delivered, missing and late counts
-alongside mean, p50, p95, p99 and maximum for capture and delivery. Warmup is
-excluded from the statistics, and a missing warmup is counted separately.
-Percentiles interpolate ranks; fewer than twenty deliveries is labelled
-`small_n`. These are descriptive sample statistics, not confidence intervals or
-evidence that the settings caused a difference. Keep endpoint, host, workload,
-binary and network conditions matched when comparing experiments, and inspect
-individual trial files before pooling.
-
-The report tool also serves as an acceptance check:
-
-```sh
-python3 scripts/audio-latency-report.py /tmp/audio-local/iter-*.json
-python3 scripts/audio-latency-report.py /tmp/audio-local/iter-*.json \
-    --format json --max-p95-ms 50 --max-missing-rate 0
-```
-
-Thresholds are experiment-specific and compare unrounded statistics; only the
-printed values are rounded. The default refuses any unsuccessful sample. A
-report with no delivered samples, or with a missing warmup, always fails,
-including when a nonzero missing-rate budget is allowed. Invalid data, an empty
-or interrupted run, a trial that left no result, failed or incomplete teardown,
-invalid timing windows, or the same run given twice fails rather than producing
-a healthy verdict; the runner then writes the refusal reason in place of
-statistics, as a JSON error object and a one-column CSV in the machine-readable
-summaries. The ordinary gate checks the report's validation, statistics, export
-formats and budget exit statuses; real-room experiments are separate from the
-credential-free gate.
-
-### Lessons from lk-latency
-
-The suite adapts the methodology of
-[lk-latency](https://github.com/haffi96/lk-latency), a LiveKit video latency
-harness: fresh rooms per iteration, raw timing archives, build identity and
-pooled statistics. It keeps CodeTrial's released SDK rather than the
-separately pinned revision that harness needs for per-stage hooks, so it does
-not invent encoder, network or decode timestamps it cannot observe. Several
-rules carry over. A run that never delivers media cannot pass. Tracks are
-unpublished before the room closes, so shutdown negotiation does not look like
-a media failure. Credentials reach the probe through the environment, never
-argv. Every timing belongs to an explicit pulse, and an ambiguous match is
-rejected rather than resolved in favor of the fastest. The endpoint is named
-for what it is, decoded PCM at the sink, with no claim about audible output.