docs(sim): scenario pack + LEARNING.md pointers (slice-4½ S8)

Three shipped scenarios assert distinct FOB reflex properties per spec section 5.3:
- loud-barge.toml: PRIMARY barge-in path (local VAD, 20 loud frames @ 20ms = 400ms)
- quiet-advisory.toml: SECONDARY path (sub-VAD-threshold quiet frames; in the
  slice-4-half standalone-wiring mode this exercises the harness's own latencies
  rather than real brain ASR-VAD -- the latter deferred to post-spearhead
  MockRealtimeBrain composition per spec section 8.6)
- sustained-call.toml: multi-barge / anti-fatigue check (3 loud bursts across 5 speak
  cycles; the second + third bar's kill-time should drift <= 1.5x the first's).

LEARNING.md gains 5 new slice-4-half concept pointers covering: measurement discipline
(the callers clock), post-hoc dedup of noise captures, in-process concurrency sweep,
atomic accumulators with compare_exchange_weak, and pub(crate) visibility for the
percentile_ms helper used by ConcurrencyRunner's sample-level aggregation (avoiding
the interleaved-captures-corrupt-LatencyProbe-pairing problem).

Signed-off-by: Aaron D. Lee <himself@adlee.work>

LEARNING.md

@@ -106,6 +106,44 @@ can read in `cargo doc --open` plus the source file itself.
   doesn't expose subprotocol/auth headers via the simple `connect_async(url)`
   entry point — see `openai_headers` / `openai_realtime_url`).
 
+## Slice 4½ — benchmark + simulation harness
+
+- **Measurement discipline: the caller's clock** →
+  `crates/rutster-sim/src/sim_audio_pipe.rs` — the `AudioPipe` test-double
+  that IS the caller. Both onset + receipt timestamps are captured inside
+  the SimAudioPipe (the wall-clock the *caller* started speaking + the
+  wall-clock the *caller* heard the reply). The harness cannot lie about
+  latency because the only clock it uses is the caller's (spec §2.2 — the
+  load-bearing design choice).
+- **Post-hoc p50/p99 computation + dedup of noise captures** →
+  `crates/rutster-sim/src/latency.rs` — the `LatencyProbe` consumes
+  `Capture` events + pairs `CallerLoudOnset` with the next
+  `BargeKillObserved` (kill-time) and the next `CallerHeardReply`
+  (mouth-to-ear). Captures without a prior onset are silently dropped
+  (the SimAudioPipe captures BargeKillObserved unconditionally on empty
+  ring; the probe is the dedup gate).
+- **In-process concurrency sweep + the doctrine-drift detector** →
+  `crates/rutster-sim/src/concurrency.rs` — the `ConcurrencyRunner` spawns
+  N `SimCall`s at levels [1, 10, 50] (the spearhead-scale envelope) +
+  aggregates per-call latencies. Per spec §2.4: 1 isolates the baseline,
+  10 is the warm working set, 50 is the saturation point.
+- **Atomic accumulators on the hot path: `compare_exchange_weak`** →
+  `crates/rutster-sim/src/tick_lag.rs` — `TickLagStats` keeps a per-tick
+  max + overrun counts via `AtomicU64`. The CAS loop on `max` is the
+  idiomatic pattern for "atomic max update" in Rust (a single
+  load-then-CAS-on-fail loop, ordering::Relaxed because stat counters
+  don't need cross-thread synchronization ordering). 3 atomic ops per tick
+  (max CAS + conditional fetch_add + unconditional fetch_add) — no Mutex,
+  lock-free hot path.
+- **`pub(crate)` visibility for cross-module helper** →
+  `crates/rutster-sim/src/latency.rs` — `percentile_ms` is `pub(crate)`
+  (not `pub` nor private). The `ConcurrencyRunner` (sibling module) needs
+  it to compute p50/p99 over the *merged sample across N probes* — merging
+  `kill_times()` samples per-probe (not concatenating `Capture` vectors)
+  avoids the interleaved-captures-corrupt-LatencyProbe-pairing problem
+  (each probe has its own `Option<Instant>` pairing cursor; merging
+  captures across probes would interleave their cursors).
+
 ## How to read
 
 1. `cargo doc --open` — every module has a `//!` doc comment; the doc

crates/rutster-sim/scenarios/loud-barge.toml

@@ -0,0 +1,27 @@
+# Scenario: loud-barge (slice-4½ spec §5.3 entry #1)
+#
+# Drives the PRIMARY barge-in path (slice-4 §5.1): the caller speaks
+# one loud burst of audio; the local VAD trips; playout dies; no brain
+# advisory needed. Asserts the wedge-#1 path: "VAD killing TTS the
+# instant the caller speaks, without the brain."
+#
+# 20 frames @ 20 ms = 400 ms of speech — comfortably past the
+# VAD_DEBOUNCE_FRAMES=3 (60 ms) debounce threshold; the VAD trips
+# deterministically.
+#
+# Threshold assertion (S7): N ∈ [1, 10, 50] concurrent SimCalls — all
+# must pass BARGE_IN_KILL_TIME_P99_MS = 80 ms at p99 (60 ms budget +
+# 20 ms observer slack).
+
+name = "loud-barge"
+
+[[steps]]
+kind = "speak_loud"
+frames = 20
+
+[[steps]]
+kind = "await_reply"
+frames = 0
+
+[[steps]]
+kind = "end"

crates/rutster-sim/scenarios/quiet-advisory.toml

@@ -0,0 +1,31 @@
+# Scenario: quiet-advisory (slice-4½ spec §5.3 entry #2)
+#
+# Drives the SECONDARY barge-in path (slice-4 §5.2): the caller speaks
+# sub-VAD-threshold audio (zeroed PCM, energy=0, well below
+# VAD_RMS_THRESHOLD=500.0); the local VAD cannot trip; the kill must
+# come from the brain's slower ASR-VAD advisory path.
+#
+# In the slice-4½ sim harness's standalone-wiring mode, the brain side
+# is a fake-brain tokio task (no real MockRealtimeBrain WS server) —
+# this scenario exercises the advisory path via the LocalVadReflex's
+# no-op observation of quiet frames (no trip) + the awaited reply from
+# the brain task's seed reply. The kill_time + mouth_to_ear metrics
+# therefore measure the harness's own latencies, not real brain-side
+# ASR-VAD latency (the latter is deferred to the post-spearhead
+# refinement tier per spec §1.2 + §8.6 — paired with MockRealtimeBrain
+# composition + LLM-driven callers).
+#
+# Threshold assertion (S7): N=1 only (the secondary-path focus).
+
+name = "quiet-advisory"
+
+[[steps]]
+kind = "speak_quiet"
+frames = 20
+
+[[steps]]
+kind = "await_reply"
+frames = 0
+
+[[steps]]
+kind = "end"

crates/rutster-sim/scenarios/sustained-call.toml

@@ -0,0 +1,49 @@
+# Scenario: sustained-call (slice-4½ spec §5.3 entry #3)
+#
+# Drives the multi-barge / fatigue / sustained-load check. Three loud
+# bursts + two quiet interludes + end:
+#   - loud cycle 1 → barge fires → local VAD re-arms on quiet
+#   - quiet cycle 1 → no barge, VAD re-armed for next loud
+#   - loud cycle 2 → second barge fires
+#   - quiet cycle 2 → re-arm
+#   - loud cycle 3 → third barge fires
+#
+# Threshold assertion (S7): the per-barge kill_times should drift
+# ≤ 1.5× across the three bar cycles (anti-fatigue). The second + third
+# bar's kill time should not be significantly longer than the first's
+# (no resource exhaustion, no GC pressure, no growing lock contention
+# across the slice's lifetime).
+#
+# See slice-4 §6.1 for the barge_epoch disambiguation mechanism that
+# makes "fresh re-barge" vs "late confirmation of the bar already in
+# flight" distinguishable — the Reflex::barge_epoch increments on each
+# SpeechStarted.
+
+name = "sustained-call"
+
+[[steps]]
+kind = "speak_loud"
+frames = 10
+
+[[steps]]
+kind = "speak_quiet"
+frames = 10
+
+[[steps]]
+kind = "speak_loud"
+frames = 10
+
+[[steps]]
+kind = "speak_quiet"
+frames = 10
+
+[[steps]]
+kind = "speak_loud"
+frames = 10
+
+[[steps]]
+kind = "await_reply"
+frames = 0
+
+[[steps]]
+kind = "end"