hsmc/docs/003. latency-analysis.md

003. Latency analysis: hsmc hot-path cost surface

This is an audit of where wall-clock and code-size cost lives in hsmc's event-dispatch path. It is intended to inform mitigation work for embedded and embassy-rs use cases where every microsecond of dispatch overhead couples directly into hardware-control loop tightness and jitter.

The audit is analytical, with three categories of evidence:

1. Source reading of hsmc/src/lib.rs, hsmc/src/journal.rs, and hsmc-macros/src/codegen.rs.
2. cargo expand --release of two test programs (one with no observation features, one with tokio + journal) to confirm what the macro really emits.
3. cargo rustc --release --emit=asm spot-checks of the compiled output to confirm LLVM-level claims (jump-table lowering, dead-code elimination of unused name tables, async-closure code size).

The audit deliberately stops short of recommending mitigations. A ranked menu of candidates is included at the end so a follow-up planning round can pick from it; nothing here is a commitment.

---

1. Cost surface, ranked by per-event impact

#	Source	Class	Per-event cost (estimate)	Location
#	Source	Class	Per-event cost (estimate)	Status
----	-----------------------------------------------	-----------------------------	------------------------------------------------------------------------------------------------------------	---------------------------------------
1	dispatch().await loop	Discretionary, async-only	2 awaits cold-start / 1 await steady-state per dispatched event. ~free if action body never Pending-yields.	open
2	LCA + transition path build	Discretionary	O(H²) heapless::Vec<u16,32>::contains scan → O(H) depth-walk → O(1) __PATH_RANGE lookup. Up/self/lateral folded into the precomputed table; lca(), is_up_transition(), and the per-call heapless::Vec<u16,16> ancestry build are all gone.	resolved
3	Parent walk for handler search	Discretionary	O(H) __PARENT walk + per-hop __handler_lookup call → O(1) __DISPATCH[(state, event)] lookup. Every (state, event) pair statically resolved at codegen. __handler_lookup retained for timer dispatch only.	resolved
4	__handler_lookup nested match (timer path)	Intrinsic	Still used for timer dispatch. LLVM lowers to jump tables (verified LJTI*). ~1 indirect jump per call.	open (already at floor)
5	run_action dispatch	Intrinsic	One match arm + monomorphized trait call + 1 vtable hop on &mut dyn QueuePush (verified devirtualized when callsite type is known).	open (already at floor)
6	Event queue push/pop	Intrinsic	heapless::Deque::push_back / pop_front, both O(1).	open (already at floor)
7	__chart_observe! macro arms	Compile-eliminable when off	Zero residue when all sink features off. Each call site collapses to {};.	verified zero-cost
8	Journal Vec::push per atom	Conditional on journal	Amortized O(1); ~16 B per TraceEvent. ~11 push sites fire per transition.	opt-in
9	Async-fn future size (binary cost)	Discretionary, async-only	~3000 asm lines of chart machinery for a 5-state chart; dominated by async-fn state machines.	open (mitigation H3)
10	tokio::sync::mpsc::unbounded_channel	Tokio-only, std-bound	Heap-backed channel for cross-task event injection. Embassy uses bounded embassy_sync::channel.	feature-specific
11	TimerTable per-step / per-exit fixed cost	Discretionary	Unconditional pop_expired + cancel_state calls every step / exit → inlined is_empty() + #[cold] non-empty body (runtime branch on charts that use timers; compile-time elided on timer-free charts).	resolved
12	TimerEntry size + per-decrement work	Discretionary	remaining_ns: u128, 32-byte entry, u128 saturating-sub per decrement → u64, 16-byte entry, u64 saturating-sub (~584-year ceiling; Duration::from_nanos round-trips exactly).	resolved
13	do_terminate() exit walk	Discretionary	Runtime __PARENT walk + pop_front-loop queue drain → precomputed __TERMINATE_DATA slice + Deque::clear() (O(1)).	resolved

Items 2, 3, 11, 12, 13 are fully resolved by the precompute / inline- empty-check passes in this round of work — see hsmc/benches/baseline.txt for the per-pass delta blocks. Items 4–6 are at or near the algorithmic floor for hierarchical state machines. Item 7 is verified zero-cost. Items 1 (async overhead) and 9 (async machinery binary size) are the only meaningful open items; both are addressable via a dispatch_sync() fast path (mitigation H3 in §9), which would be a user-facing API change and is therefore out-of-scope for this round.

---

2. Per-event cost model (analytical)

For a chart with hierarchy depth H=3, one entry action + one exit action, a simple lateral transition, no observation features, cargo build --release, on a Cortex-M4-class core (~168 MHz):

Stage	Cycles (rough)	Notes
queue.pop_front	~5–10	heapless::Deque index op
__event_to_id	~3–5	match arm (jump table)
__DISPATCH[(state, event)] lookup	~5	one indexed load + sentinel branch
run_action (1 action)	~10–20	match + monomorphized call + 1 vtable hop
Path lookup (__PATH_RANGE)	~5	one indexed load + slice arithmetic
Exit chain (H states × A actions)	~3 × H + 30 × A	per-state work is now ~3 instr each
Enter chain (H states × A actions)	~3 × H + 30 × A	per-state work is now ~3 instr each
Timer machinery (timer-free chart)	0	codegen-elided when ir.duration_triggers.is_empty()
Timer machinery (timer chart, no fire)	~3	inlined is_empty() branches × 3
Timer machinery (timer firing)	~30–40	decrement + pop_expired + dispatch_trigger
Total .step() steady-state	~0.4 – 0.8 µs	for H=3, A=1; user action body excluded
Async overhead (.dispatch() loop)	per-await checkpoint	~free if no Pending-yield

These are back-of-envelope numbers; actual measurements would tighten them by ~2× in either direction, depending on inlining decisions.

For empirical numbers (deterministic, instruction-counted), see hsmc/benches/baseline.txt. Headline numbers (current): h2_warm_lateral 84 instr / 397 cycles; h4_cross_tree 101 instr / 456 cycles; h8_cross_tree 93 instr / 443 cycles; bubble_top 99 instr / 453 cycles (4-hop handler resolution); bubble_cross_tree 94 instr / 444 cycles (4-hop bubble + 10-state cross-tree); terminate_event 69 instr / 311 cycles; timer_fire_dispatch 191 instr / 706 cycles; chain_self_emit 419 instr across 4 events (≈105 instr/event). Cumulative on h8_cross_tree: -82% instructions / -72% cycles vs the original O(H²) LCA scan.

---

3. Spot-check evidence

3.1 __chart_observe! is zero-cost when off (cargo expand)

Test: tests/state_chart_const.rs, with no observation features enabled.

In the expanded source (/tmp/hsmc_expand_no_features.rs, 911 lines), every observation site collapses to an empty block:

// in `step()`:
if self.current.is_none() {
    {};                     // would be observe_started
    self.enter_path(0);
    return self.timers.min_remaining();
}
// ...
if let Some((state, trig)) = self.timers.pop_expired(&__DEPTH) {
    let __t_state = state;
    let __t_trigger = trig;
    {};                     // would be observe_timer_fired
    // ...
}

A grep for any of TraceEvent::, ::log::, ::tracing::, ::defmt::, JournalSink::push in the expanded code returns zero matches. The macro arms entirely vanish.

JournalSink itself becomes a unit struct when the feature is off:

#[cfg(not(feature = "journal"))]
pub struct JournalSink;     // ZST — 0 bytes in the chart struct

so the __journal field on the chart costs zero bytes at runtime.

3.2 Name lookups are dead-code-eliminated (cargo rustc --emit=asm)

Test: tests/det_lateral.rs, built with --features tokio,journal.

Even with journal enabled, a grep across the 22,477 lines of generated asm for state_name, action_name, event_name, or timer_name returns zero matches. The journal stores u16 indices, not strings, and no trace-* feature is on, so the name-lookup helper functions are uncalled and DCE'd from the final binary.

This means even if a user enables journal for replay/verification, they pay zero binary cost for the human-readable name tables until they also enable a trace-* sink.

3.3 LLVM lowers handler matches to jump tables

Same asm file. Grep for LJTI:

    leaq    .LJTI2_0(%rip), %rcx
    jmpq    *%rax
.LJTI2_0:
    .long   .LBB2_1-.LJTI2_0
    .long   .LBB2_28-.LJTI2_0
    .long   .LBB2_29-.LJTI2_0
    .long   .LBB2_23-.LJTI2_0

Multiple jump tables are present, each sized roughly to the chart's state/event count. This corroborates the assumption that __handler_lookup's nested match (state_id) { match (kind, trigger_id) } lowers to a single indirect jump on platforms LLVM can target with one, not a chain of compares.

3.4 Async machinery code size

Symbol-by-symbol asm sizes from det_lateral:

Symbol (async closure / future)	Asm lines	Notes
Lat::exit_state	~275	one .await per exit action
Lat::run_action	~349	match arm + N monomorphized async branches
Lat::run_handlers	~826	loop over actions, each .await
Lat::enter_state_no_descent	~217
Lat::step	~970	inlines lca, is_up_transition, etc.
Lat::dispatch	~137	thin loop wrapper
Lat-chart total	~2,800	async machinery for one 5-state chart

For comparison, Lat::new is 141 lines. The async coloring roughly doubles to triples the binary footprint of a chart relative to a pure sync codegen. On Cortex-M0/M0+ where flash is small, this matters.

3.5 Journal expansion adds N pushes, nothing else (cargo expand)

Same test, expanded with --features tokio,journal (/tmp/hsmc_expand_with_journal.rs, 2,426 lines vs. 911 lines without).

Counts:

• 104 total TraceEvent::* constructions.
• 20 distinct variant names (the full vocabulary).
• 8 Vec::push call sites inside step() — but only one fires per invocation (timer-fired path, event-received path, etc., are mutually exclusive).

The diff against the no-feature expansion is exactly: every {}; placeholder becomes a (&mut self.__journal).push(TraceEvent::...). No other infrastructure is added. No locking, no allocation per push beyond the amortized Vec growth.

---

4. Observability cost matrix

Distilled from lib.rs:158-712 and the spot-checks above:

Feature(s)	Per-call CPU	Per-call alloc	Off → fully elided?
(none)	0	0	Yes (verified)
journal	Vec::push ~10 ns amortized	~16-24 B amortized	n/a
trace-defmt	format-hash u32 + field write	0	Yes
trace-log	log_enabled!() check + format if passes	0 unless format	Yes (with small residual when on but level filtered)
trace-tracing	unconditional field collection + span	0	Yes (with ~30-cycle residual when on but no subscriber)

The codegen emits ~11 __chart_observe! calls per typical transition in deterministic order: EventReceived → EventDelivered → TransitionFired → ExitBegan → ActionInvoked{Exit} → Exited → EnterBegan → ActionInvoked{Entry} → Entered → TransitionComplete, plus action-handler markers. That's the multiplier on every per-call number above.

---

5. What is not a problem (audited and dismissed)

• Handler-lookup growth with chart size — confirmed jump-table lowering (§3.3). O(1) per event regardless of state or event count (modulo cache effects from larger jump tables).
• Trait-object cost on action dispatch — only the &mut dyn QueuePush proxy is dyn. User actions are monomorphized trait methods (one match arm per action ID, direct calls). One vtable hop per emit, not per action.
• Observation overhead with features off — confirmed zero. Verified three ways: macro arm structure (every body #[cfg]-gated), expanded source (every site is {};), and asm (no name-lookup symbols, no trace symbols).
• Heap allocation in default build — extern crate alloc is confined to journal.rs:26. With journal off, the runtime is fully stack-only. Verified: no Vec, Box, or String in expanded sync code.
• Pin<Box<...>> futures on the hot path — none. Chart futures are inline state machines, not boxed.
• Tokio unbounded_channel for embedded — irrelevant: tokio is std anyway. Embassy uses bounded embassy_sync::channel::Channel (no alloc).

---

6. Event entry points

Events reach the chart's internal heapless::Deque via four distinct paths. Each has its own latency, wakeup, ordering, and backpressure profile, summarized below.

Path	API	Backing	Wakeup	Ordering	Backpressure
1	chart.send(event) (sync, single-task)	heapless::Deque::push_back	none — caller drives	FIFO within queue	Err(QueueFull) immediate
2a	chart.sender().send(ev) (tokio)	tokio::sync::mpsc::unbounded_channel	tokio task wake	FIFO; biased select deprioritizes vs. durings	unbounded memory
2b	chart.sender().send(ev) (embassy)	embassy_sync::channel::Channel (bounded)	embassy task wake	FIFO; same biased-select	Err(Full) (try_send) or backpressure-await
3	actx.emit(ev) from inside an action	same internal Deque as path 1	none — drained inline	drained synchronously after current handler completes; before step() returns	sticky overflow flag
4	return from a during: future	select! arm in run loop	driver-specific	biased priority over channel events	silent drop on Deque full

Path 1 (sync send) — codegen.rs:2316-2319. Bare Deque::push_back; owning task only, no synchronization. ~5–10 cycles. Caller is responsible for invoking step() (or dispatch() under an async feature) to actually process the event.

Path 2 (cross-task sender()) — codegen.rs:1492 (tokio sender), 1638 (embassy sender), 1308 / 1405 (run-loop receive sites). The canonical cross-thread / ISR path per the spec. Tokio uses an unbounded MPSC (heap-backed; no backpressure signal); embassy uses a bounded Channel<M, E, N> with try_send for ISR-safe non-blocking emits and an async send that backpressures the producer. The receiver task wakes on send; one iteration of the run-loop's select! consumes the event and pushes it into the internal Deque.

Path 3 (self-emit from action) — actx.emit(...) goes through EmitProxy::push (hsmc/src/lib.rs) which writes into the same internal Deque as Path 1. Self-emitted events are drained in the same step() call that emitted them — dispatch() runs while !queue.is_empty() { step().await } until quiescence. Nothing crosses a task boundary.

Path 4 (during: return) — codegen.rs:1306–1313 (tokio with durings), 1392+ (embassy). Durings are racing futures spawned in the run-loop's select!. When one resolves to an event, the select! unblocks and the event is pushed into the internal Deque. The select! is biased; so durings have priority over the channel — relevant if a during is a poll-based driver that's almost always ready (see §8).

One ordering caveat: there is no global FIFO across paths 1–4. Within each path, events are FIFO; across paths, the order depends on when each path's wakeup arrives at the run loop. For most charts this is a non-issue, but a chart that mixes high-frequency cross-task sends with rapid self-emits could observe self-emits processed before external events that arrived "first" in wall-clock terms.

---

7. Cooperative-driving audit

The library targets embassy-rs's single-priority cooperative scheduler, where any tight loop without an .await can starve other tasks. An audit of every loop and await pattern in the runtime and codegen:

Loop / await	Location	Bounded?	Awaits per iter?	Verdict
Main run() outer	codegen.rs:1290+, 1379+	by terminated flag	yes — always	cooperative
dispatch() drain loop	codegen.rs:2339	by queue size	only via step().await	suspect — fairness
Transition exit/enter slice loops	codegen.rs (transition)	by precomputed slice len	yes (async variant)	OK
Termination drain	codegen.rs:2750	by hierarchy + queue	yes (async variant)	OK
Timer decrement / pop	runtime / lib.rs	by timer-table size	sync, off the event-loop	OK
Send/recv backends	codegen.rs:1308, 1405	n/a	always .await	cooperative

No try_recv patterns anywhere. No fixed-period polling. The library is fundamentally event-driven and embassy-friendly.

One latent fairness characteristic: the dispatch() drain loop — while !self.queue.is_empty() { step(Duration::ZERO).await } — never yields to the executor between iterations. If user actions are all synchronous and they emit a chain of follow-on events (action A emits B, B emits C, …), each step().await completes inline, and the loop becomes effectively synchronous for the duration of the chain. Embassy high-priority tasks could be delayed by the chain length.

This is not a bug — dispatch() is documented to drain to quiescence — but it's worth knowing about. A future option (not part of this round of work) would be a feature-gated embassy_futures::yield_now().await every Nth iteration, or a chart-level config to bound chain length. Not done by default because it would serialize the common case for no benefit.

---

8. Correctness invariants surfaced during the audit

Three contracts that aren't bugs, but warrant explicit documentation because they are easy to miss when reading the code:

Sticky overflow flag in EmitProxy (hsmc/src/lib.rs, EmitProxy::push). When a self-emit fails because the internal Deque is full, an __overflow flag is set on the chart. Every subsequent actx.emit(...) in the same dispatch chain returns Err(QueueFull), even after the queue has drained. The flag is cleared at the top of each dispatch() call by take_overflow(). This is intentional — it prevents a chart from observing a partially-applied sequence of self-emissions where some made it into the queue and others did not. Authors of actions that emit multiple events should treat Err(QueueFull) as terminal for the current chain.

Biased select in run loop (tokio: codegen.rs:1306–1313; embassy: 1392+). The select! is biased;, so durings are always polled before the cross-task channel. For most charts this is the right priority — a sensor driver should pre-empt a UI event. But for a during: that wraps a poll-based driver (one that returns Ready nearly every poll), the channel can be starved indefinitely. Cure: make the during yield (embassy_futures::yield_now) when there's nothing real to report.

Cancel-safety contract on durings (hsmc/src/lib.rs:53–60). Every .await in a during: future must be a clean resume point, because the future will be dropped if the state transitions before it completes. This is the standard async-cancel-safety contract, but it is easy to violate by holding a mutable borrow across an await — the borrow is "valid" when resumed but the surrounding state may have moved on.

---

9. Candidate mitigations (menu, NOT a commitment)

Sized for sequencing in a follow-up plan-mode session. "Effort" is gut feel: L = afternoon, M = couple of days, H = a week or more.

ID	Mitigation	Targets	Impact	Effort	Cost
H0	Replace lca() with an O(H) depth-walk via the __DEPTH table.	#2	Landed.	-	-
H1	Codegen per-(from, to) precomputed exit/enter chains as __PATH_DATA + __PATH_RANGE.	#2	Landed. Up/self/lateral folded into one indexed lookup; lca() and is_up_transition() removed.	-	~6·N² B index + ~N²·H B path data
T1	Inlined is_empty() early-out on pop_expired / min_remaining / decrement / cancel_state with #[cold] non-empty bodies.	#11	Landed.	-	-
T2	Codegen-static elide of all TimerTable calls when ir.duration_triggers.is_empty().	#11	Landed. Most charts are timer-free; step() returns None directly.	-	-
H4	__DISPATCH[(state, event)] const array; one indexed lookup replaces the runtime __PARENT bubble walk + per-hop __handler_lookup.	#3	Landed. Each (state, event) statically resolved at codegen.	-	~24 B per (state, event) pair
U1	Shrink TimerEntry.remaining_ns: u128 → u64 (~584-year ceiling). Halves entry size; replaces u128 saturating-sub with u64.	#12	Landed.	-	-
K1	__TERMINATE_DATA per-state precomputed exit chain (root inclusive); do_terminate() indexes a slice instead of walking __PARENT, and uses Deque::clear() (O(1)) for the queue drain.	#13	Landed. terminate_event -22.5% instr / -23.4% cycles.	-	~(H+1)·N B path data
H3	Provide a dispatch_sync() API that bypasses .await when the chart's actions are fully synchronous.	#1, #9	High for sync-action charts	L-M	API surface growth (user-facing)
H5	Sparse __PATH_RANGE / __DISPATCH for very large charts (N > 64): emit only reachable pairs with a small perfect-hash. Latency-neutral; cuts O(N²) static memory growth.	cost mitigation	Low at typical N; significant at N > 100	M	codegen complexity
M1	Eliminate the trailing step().await in dispatch() when queue empties mid-step.	#1	Low	L	minor refactor
M2	Lazy field eval in trace-tracing arm using tracing's field-closure form to skip evaluation when filtered.	#8 (on)	Low	L	none

H0/H1/T1/T2/H4/U1/K1 all landed in the most recent round of work. Callgrind now reports the entire HSM machinery as a single step symbol — every helper inlined. The only remaining optimization avenue that meaningfully affects per-event latency is H3 (dispatch_sync()), which is excluded as user-facing. Memory-side mitigations (H5) only matter for N > 64-state charts; embedded users with large charts are the audience.

---

10. Where to look in the source

For anyone reading this audit and wanting to verify or extend it:

• hsmc/src/lib.rs
  • :94 — the #![cfg_attr(... no_std)] boundary
  • :158-712 — __chart_observe! macro, all observation arms
  • :737-772 — JournalSink ZST/Vec duality
• hsmc/src/journal.rs
  • :26-32 — extern crate alloc and Vec import (only here)
  • :65-150+ — TraceEvent enum (the journal vocabulary)
• hsmc-macros/src/codegen.rs
  • compute_transition_paths() (top of file, after compute_chart_hash) — builds __PATH_DATA + __PATH_RANGE at codegen time
  • __PARENT, __DEPTH, __DEFAULT_CHILD, __PATH_DATA, __PATH_RANGE static-table emission (search "const __PARENT")
  • step() (search "fn step" / "async fn step") — sync vs. async branches
  • dispatch() await loop (search "fn dispatch")
  • run_action (search "fn run_action") — sync vs. async
  • transition (search "fn transition") — now a thin slice-iter wrapper
  • name-lookup helpers (__state_name, __action_name, etc.) at :631-667
• Test reference shape: hsmc/tests/state_chart_const.rs is the smallest sync-only chart in the tree; hsmc/tests/det_lateral.rs is the reference async + journal shape used for the asm spot-check.

---

11. Reproducing the spot-checks

# Zero-cost observation verification (expand → no observation residue)
cargo expand --release --test state_chart_const \
    > /tmp/expand_no_features.rs --manifest-path hsmc/Cargo.toml
grep -E "TraceEvent::|::log::|::tracing::|::defmt::" /tmp/expand_no_features.rs
# expect: no matches

# Journal expansion (count of TraceEvent push sites)
cargo expand --release --test det_lateral --features tokio,journal \
    --manifest-path hsmc/Cargo.toml \
    | grep -c "TraceEvent::"
# expect: ~100

# Asm spot-check (jump tables, function sizes, name DCE)
cargo rustc --release --test det_lateral --features tokio,journal \
    --manifest-path hsmc/Cargo.toml -- --emit=asm
ASM=$(find target -name 'det_lateral*.s' | head -1)
grep -E "LJTI|jmpq?[[:space:]]+\*" "$ASM"        # jump tables
grep -E "state_name|action_name" "$ASM"           # expect no matches

---

12. Summary

hsmc's hot path is mostly already at the algorithmic floor for hierarchical state machines. The intrinsic per-event cost (parent walk, handler lookup, action dispatch, queue ops) is small and well-monomorphized. Observation cost is verified zero when features are off, and stays proportional to the user's chosen sink mix when on.

Items #2, #3, #11, #12, and #13 in §1 have all landed across six optimization passes (H0 → H1 → T1 → H4 → T2 → K1+U1). Cumulative impact on the deepest dispatch (h8_cross_tree) is -82% instructions / -72% cycles vs the original O(H²) LCA scan. The headline wins by bench:

• terminate_event 69 instr / 311 cycles (~14 instr of HSM work after subtracting bench wrapper) — do_terminate() is one indexed slice
  • Deque::clear().
• bubble_top 99 instr / 453 cycles — 4-hop handler resolution collapses to one indexed __DISPATCH lookup; per-dispatch cost is bench-shape-independent.
• bubble_cross_tree 94 instr / 444 cycles — realistic HSM worst case (deep bubble + 10-state cross-tree) costs the same as a plain cross-tree because bubble lookup is O(1) and transition is O(path-length).
• h8_cross_tree 93 instr / 443 cycles — 7-deep cross-tree dispatch.

Callgrind reports the entire HSM machinery as a single step symbol — run_handlers, run_action, transition, do_terminate, exit_state, enter_state_no_descent, EmitProxy::push, all four TimerTable methods, the dispatch-table lookup, and the terminate-path lookup are all inlined.

Cost remaining, in priority order:

1. Async coloring code-size cost (item #9) — chart machinery is ~3,000 asm lines for a 5-state chart, dominated by async-fn state machines. Mitigation H3 (a dispatch_sync() fast path) would let users with all-sync actions skip this entirely. User-facing API change — out of scope for the current optimization pass.
2. Bench-wrapper noise floor — at the smallest dispatch's ~38–80 instructions, the iai-callgrind wrapper (black_box, send, instrumented call-edge crossing) accounts for ~45 instructions. Sub-bench-wrapper improvements are below the measurement floor.
3. Memory cost of __PATH_DATA / __PATH_RANGE / __DISPATCH / __TERMINATE_DATA at large N — __PATH_RANGE is O(N²); __DISPATCH is O(N×E); __TERMINATE_DATA is O(N×H). For N ≤ 32 / E ≤ 16 / H ≤ 8 the total static footprint is a few KiB. A 64-state × 32-event chart would push past ~80 KiB. Mitigation H5 sketches a sparse encoding.

All addressable purely in hsmc-macros/src/codegen.rs and hsmc/src/lib.rs without breaking runtime semantics or feature gates. The bench harness in hsmc/benches/dispatch.rs (13 benches, every dispatch shape covered) is in place to measure each mitigation as it lands.