lading-optimize-find-target

Phase 1: Discover Benchmark-Eligible Modules

A module is eligible if it has a Criterion benchmark in lading_payload/benches/.

1. List benchmarks: Glob lading_payload/benches/*.rs — each filename (minus .rs) is a bench name
2. Resolve sources: For each bench name, find the corresponding source file(s) under lading_payload/src/. Check single-file modules ({name}.rs), directory modules ({name}/), and parent-module patterns (e.g., opentelemetry_log → opentelemetry/log.rs)
3. Match fingerprints: Glob ci/fingerprints/*/lading.yaml — each directory name is a fingerprint. Match fingerprints to modules by reading the variant: key from each config

The result is a set of (bench, source_files, fingerprint_or_none) triples.

---

Phase 2: Profile Allocation Intensity

Run the profiling script:

.claude/skills/lading-optimize-find-target/scripts/profile-modules

Output is TSV: module, allocations, total_bytes, peak_live_bytes. Record per-module.

---

Phase 3: Learn from Past Optimizations

Read the optimization history to understand what techniques work and what's already done:

Read .claude/skills/lading-optimize-hunt/assets/db.yaml

For each entry, read its detail file (file: field, relative to .claude/skills/lading-optimize-hunt/) to extract:

• Technique and measurements — which techniques yielded what % improvements
• Lessons — what patterns were optimized and what the before/after looked like
• Targets already covered — so you skip them

This history teaches you what to look for. Successful past techniques are strong signals for where to look next. The lessons field often suggests next targets explicitly.

---

Phase 4: Find Opportunities

Scan every benchmark-eligible source module for every known pattern below. This is an exhaustive cross-product — do not short-circuit after finding one hit.

Known Patterns

Name	Pattern	Technique
vec-with-capacity	Vec::new() + repeated push	Vec::with_capacity(n)
string-with-capacity	String::new() + repeated push	String::with_capacity(n)
map-with-capacity	FxHashMap::default() hot insert	FxHashMap::with_capacity(n)
buffer-reuse	format!() in hot loop	write!() to reused buffer
slice-params	&Vec<T> or &String parameter	&[T] or &str slice
hoist-allocation	Allocation in hot loop	Move allocation outside loop
object-pool	Repeated temp allocations	Object pool / buffer reuse
borrow-not-clone	Clone where borrow works	Use reference
inline	Hot cross-crate fn call	#[inline] attribute
lazy-iterators	Intermediate .collect() calls	Iterator chains without collect
box-large-structs	Large struct by value	Box or reference
bounded-buffer	Unbounded growth	Bounded buffer with .clear()
scratch-buffer	encode_to_vec() per call	Reusable BytesMut scratch buffer
on-demand-serialization	Deep clone of template in loop	Incremental mutation / COW