Profilesmash

Audit the instruments, not the program. performancesmash asks where the time goes; this asks whether anyone could find out. A project without attribution optimizes by anecdote, and one with untrustworthy numbers optimizes the wrong thing with confidence. Agents make it worse: they add a benchmark that compiles to nothing, report a mean from one run on a laptop, and call it measured.

Audit only. Do not edit, commit, or push. You may build, run benchmarks, and run profilers locally, writing output to the scratchpad. A slowdown probe (inject a sleep, an extra allocation, or an extra query, then run the gates) is allowed only in a throwaway worktree or copy. Never generate load against shared, staging, or production systems, and never attach to a production process.

You are the coordinator. Continue until every asset has a verdict, every matrix cell is filled, and the final audit is validated.

The Three Questions

Ask them of every critical flow, for every cost dimension it has:

1. Can someone attribute it? Given "this got slow" or "this uses too much", can a person or an agent get from the symptom to a line, query, span, or allocation site with one documented command, in minutes? A total is not an answer.
2. Can the number be trusted? Production build and flags, production-shaped workload, a distribution instead of a mean, known noise, known overhead, symbols that resolve.
3. Would a regression be noticed? By a gate before merge, a trend after it, an alert in production, or nobody.

Dimensions

Each is a different question needing a different instrument. A project that owns only a CPU flame graph diagnoses everything as CPU.