scripts: add bench-dedupe.py, a larger-than-RAM dedupe benchmark harness

[martinus]

What

Commits the dedupe-phase benchmark harness that keeps getting reconstructed by hand for each release. scripts/bench.py deliberately only measures the scan/hash phase (-rq, non-destructive); it can't measure dedupe — dedupe mutates the tree (each round needs a fresh unshared copy) and its signature cost only shows up when the working set is larger than RAM. scripts/bench-dedupe.py fills that gap and produces the numbers in docs/benchmarks.md's Larger-than-RAM dedupe section.

What it does

• Builds N --reflink=never copies of a --source tree (dedupe reclaims the later copies against copy0).
• Times two phases per binary — hash (-rq, in-memory) and dedupe (-dr --hashfile) — cold (drop_caches) inside a MemoryMax-capped systemd-run --user --scope, interleaved in randomized order, restoring the copies + dropping the hashfile between dedupe rounds.
• Reports wall/user/sys/RSS/disk-read as median/mean/min/max (+ hashfile size).
• --verify confirms byte-identical sharing across binaries via btrfs filesystem du -s.
• --baseline build:REF builds a comparison binary from a git ref in a cached worktree (build:897a222 = the fork's pure-upstream base, the only valid duperemove baseline); or pass a path, or omit it to bench oans alone.
• Reuses bench.py's drop_caches / reflink-fs helpers rather than duplicating them.

Canonical A/B:

scripts/bench-dedupe.py --baseline build:897a222 --source ~/git/linux --cap 4G --rounds 10 --verify

Verified

• py_compile + --help clean.
• End-to-end smoke (2 copies of a small tree, both phases, --verify, capped scope, --clean): tables render, hashfile sizes captured, verify reports IDENTICAL sharing, cleanup works.
• The build:897a222 baseline path: adds the worktree, builds it, runs, and --clean removes the worktree.

Dev tooling + docs only — no source changes. docs/benchmarks.md "Reproduce it" and CLAUDE.md now point at the tool.

🤖 Generated with Claude Code