[m-hu]

Fixes #1. Branch: m-hu:perf/1-callgrind. Spec: docs/1/specification.md (on this branch).

Problems

• The algorithms repo has no performance signal: JVM-vs-clojure-rs
  comparisons do not exist here, while the snippet repo has had a
  deterministic callgrind sweep with Bencher overlays since T1211.
• Four test-isolation bugs hid in the suite, invisible to the
  whole-suite runners: three namespaces use clojure.set /
  clojure.java.io without requiring them (passing only by ambient
  load order), and one test depends on stack exhaustion
  ((thrown? Throwable (power 2 -1)) with an impl that recurses
  past zero forever).

Root causes

• No sweep tooling or workflow exists in this repo (ci.yml runs
  correctness only).
• Per-namespace standalone runs (required for per-benchmark
  measurement) expose load-order dependence: karger-test,
  xor-cipher-test, and disjoint-set-test fail with
  ClassNotFoundException / Syntax error compiling on a fresh
  runtime on both legs (verified JVM fails identically — repo
  bugs, not port gaps).
• (power 2 -1) never terminates by design; the JVM surfaces
  this as a catchable StackOverflowError, while clojure-rs
  aborts the process (SIGABRT, exit 134 — port gap tracked at
  Unbounded recursion aborts the process instead of throwing a catchable StackOverflowError clojure#220).

Solutions (why each resolves its problem)

• perf/driver.clj: one driver for both legs — require the ns,
  clojure.test/run-tests, print one
  SUMMARY {:test :pass :fail :error} line. Same workload both
  legs; no System/exit (unemulated natively) — the gate reads
  stdout.
• scripts/ci/go (algoperf): vendored recipe-identical measure
  core (valgrind flags, JVM -Xint+EpsilonGC, LD_PRELOAD shim
  with byte-identical C source, BMF shape) with namespace
  discovery and the SUMMARY gate instead of full-stdout
  comparison (namespaces print nondeterministically).
• perf-callgrind.yml: resolve-rs (floats clojure-rs HEAD) +
  correctness gate + shard(8) x runtime(2) sweep + per-runtime
  merge + JVM-vs-native count-diff gate. ci.yml untouched.
• Test-isolation fixes (spec §non-goals amended — the sweep
  itself surfaced them): three missing :requires, one
  fail-fast negative-exponent guard.
• JVM locale pin (-Duser.language/country): comma-decimal
  hosts failed 13 namespaces via format-based assertions and
  would perturb Ir counts.
• Split timeouts (--timeout 1800 native / --callgrind-timeout
  600 + skip-degrade): the slowest namespace needs ~11min on
  the native leg, which must not flake the correctness gate.

No regressions

• ci.yml, deps.edn, build.clj untouched; the sweep is a new
  workflow file plus additive perf/ + scripts/ + docs/.
• The four src//test/ fixes keep full-suite semantics: the
  requires are no-ops where already loaded, and the power guard
  throws a Throwable exactly where the test asserts one
  (whole-suite clojure -M:test still the gate's precondition).
• Recipe fidelity: the C shim is md5-identical to T1211's, and
  JVM flags/parsers/BMF are line ports (only discovery, gating,
  and naming adapted — all covered by self-tests).

Test Plan

• gofmt -l clean, go vet ./..., go test ./... green
  (cmd/algoperf, internal/algocore, internal/measure).
• algoperf measure --self-test: 26/26 checks (SUMMARY,
  callgrind, shim, EpsilonGC, GNU-time parsers; templates;
  leg/skip classifiers).
• python3 -c yaml.safe_load on the workflow + action;
  git diff --check clean.
• Native e2e (--only, 3 namespaces): gate passes, report +
  BMF + summary outputs verified by shape.
• JVM e2e incl. EpsilonGC alloc path (alloc_min=59430912B
  on one namespace).
• --merge verified (dedupe + counts + summary render).
• Full 1351-namespace sweeps, timed phase, both legs:
  1351 ok / 0 failed each; cross-leg counts match except
  the 2 documented random-graph namespaces (karger,
  random-graph-generator), whose :pass variance is
  normalized in report-compare (proven the only varying
  ones over the full 1351x2 data).
• CI (Linux) remains the proof for the valgrind/shim paths
  (no valgrind on this host) — same as T1211.

Housekeeping notes

• Branch commits predate the AI-Tool: trailer convention on
  this repo (pushed; not rewritten per policy) — attribution
  rides this PR body into the squash instead. New commits on
  this branch will carry the trailer.
• Draft until the admin step lands (below).

Admin needed (repo admin)

• Create Bencher project clojure-algorithms, set
  BENCHER_API_KEY secret (Bencher steps fail without it).
• Optionally require the new checks in branch protection
  after first green.

Session: gentle-menkar.

AI-Tool: Muse Code 1.4.14-R7309.1 (Meta Muse Spark)

[m-hu]

Round-2 push (4c2d268) from CI round-1 evidence (all 8 native shard artifacts harvested; JVM legs still running then):

• perf/sweep_skips.json excludes matrix-exponentiation-test with a reviewed reason: native timed run >1800s on CI, JVM callgrind OOMs the 2g Epsilon heap (reproduced locally). Skips are recorded in the report, never silent.
• Shim cross-check now compares :test/:fail/:error + :pass>0 — the 2 random-graph namespaces flaked the exact tally comparison.
• JVM callgrind OOM reports the heap hint instead of cryptic exit 3.

Round-1 native sweep itself: 1350/1351 ok with full callgrind data (167/169 per shard typical), 7 callgrind timeouts degrading to skips as designed, 1 transient missing shim marker (cheap-progress, single occurrence — watching for recurrence). Remaining red is expected to be Bencher auth only.

Session: gentle-menkar.

AI-Tool: Muse Code 1.4.14-R7309.1 (Meta Muse Spark)

[m-hu]

Round-3 push (15824e1): round-2 post-mortem complete — 15/16 legs fail ONLY on the Bencher step (missing secret, expected); the one real sweep failure is JVM shard 3's problem-056.sol1-test, which OOMs the 2g EpsilonGC heap (measured 7.4g locally; next-heaviest namespace is 0.67g, so the budget stays and the namespace joins sweep_skips.json with the measured reason).

Round-2 sweep health from artifacts: all native shards 0 sweep failures with full callgrind data; all JVM shards 0 sweep failures except problem-056.

Session: gentle-menkar.

AI-Tool: Muse Code 1.4.14-R7309.1 (Meta Muse Spark)

[m-hu]

Round-3 terminal state (15824e1, run 41579900347): all 16 legs fail ONLY on the Bencher step (missing BENCHER_API_KEY — the known admin blocker). Everything else validated, including the report jobs' exact commands replayed locally against all 16 round-3 artifacts:

• Merge: 1351 results -> 1349 benchmarks (2 skips), 0 failures, both legs; matches jq -s add union on both legs.
• Count-diff: JVM vs native test counts MATCH with the 2 documented random-graph exceptions.
• Callgrind coverage: JVM 1342/1349 benchmarks with both measures (7 timeout-skips by design), native 1341/1349 (8 skips).

No code changes needed. Remaining: create the Bencher project + secret (admin checklist on this PR), then re-run to green and land.

Session: gentle-menkar.

AI-Tool: Muse Code 1.4.14-R7309.1 (Meta Muse Spark)

[m-hu]

Admin step landed: BENCHER_API_KEY secret set, and the token + clojure-algorithms project verified with a probe submission (report on throwaway branch ci-probe). Re-ran the 16 failed legs of run 41579900347 — Bencher steps should report now, then merge + count-diff jobs run.

Session: gentle-menkar.

AI-Tool: Muse Code 1.4.14-R7309.1 (Meta Muse Spark)

[m-hu]

Session: gentle-menkar. Closing as superseded by #3.

Root cause of the red Bencher steps (all rounds, including after the secret existed): this PR head is on the m-hu fork, and pull_request-triggered runs from forks receive no secrets — so BENCHER_API_KEY always rendered as empty and bencher run died on --key <KEY> validation. The sweeps themselves were green in every round (all 16 shards uploaded results); only the report step failed.

Fix: same branch content pushed to a same-repo head (hhh-pa/clojure-algorithms:perf/1-callgrind), where secrets flow to pull_request runs. Continuing in #3.