[m14] kineto profiler receiver: RFC-0012 + pkg/kineto/ parser + receiver + bench scaffold

[trilamsr]

Summary

Ships M14 (Kineto profiler receiver) end-to-end across the four phases of the M14 plan (plan is gitignored locally; design lives in RFC-0012 and the spec). Branch is one logical PR per autonomous-run constraint; split into 4 cuts before merge if preferred.

• RFC-0012 design-locks the receiver. Originally drafted as RFC-0010 but renumbered after [docs] rfc-0010: containerstdout receiver scope (M15, alpha; design-locked draft) #94 and an in-flight M16 RFC claimed the slot.
• MILESTONES.md M14 rubric corrections: step-id source (ProfilerStep#N interval, not args.Iteration — falsified during design review against PyTorch source); sampling determinism (per-(rank, step) repeated-lookup equality, not cross-rank set equality; 10000 mod granularity); fr_trace analogy disclaimer.
• pkg/kineto/ streaming Chrome-trace parser. stdlib-only, no internal/ imports. Typed sentinel errors, Synthesize deterministic generator, toy_2step checked-in fixture + golden + SHA256SUMS, FuzzParseKinetoTrace 30s in-tree, BenchmarkParse_50MB advisory. make generate-fixtures-check wired into make ci.
• components/receivers/kineto/ receiver: single-pass ProfilerStep#N stack + complete-phase interval tracker, /proc/<pid>/environ rank discovery (with filename-fallback for tensorboard_trace_handler worker prefix; with receiver-env fallback), deterministic per-(rank, step) sampling, optional cpu_op aggregation within a step, fsnotify IN_CLOSE_WRITE trigger (gz-aware glob), safe.Call wrap, degraded-state machine, README + RUNBOOK.
• Factory registered in components.yaml; cmd/tracecore/components.go regenerated.
• MILESTONES.md M14 flipped ☐ → ☑ alpha; functional rubrics ☑, non-functional ⧗ pending PR D wall-time runs.
• Bench scaffold: bench/overhead/kineto_bench_test.go (10-min CPU gate), soak_test.go (//go:build soak, 100 × 50MB RSS gate), stress_2gb_test.go (//go:build !race, 2GB HeapAlloc ceiling). Wall-time runs deferred — code lands, ops can flip rubrics once gates run.

Verified during design (May 2026)

Cited verbatim in RFC-0012:

• pytorch/pytorch torch/profiler/profiler.py:44, 651, 655, 1165-1169, 1230-1233 — PROFILER_STEP_NAME, default filename {worker}.{ts_ns}.pt.trace.json, direct-write export_chrome_trace (no tmp+rename — IN_CLOSE_WRITE is the safe trigger), record_function(\"ProfilerStep#N\").
• pytorch/kineto libkineto/src/output_json.cpp:341-352, 543, 557, 580, 593 — top-level key order; traceEvents written between handleTraceStart and finalizeTrace; no top-level samples or stackFrames (streaming-from-the-top decoder is safe).

Test plan

• `go test -race ./pkg/kineto/...` — PASS
• `go test -race -short ./components/receivers/kineto/...` — PASS (all sub-tests: config, step tracker, rank, emit, sampling, ingest golden, aggregation, factory, shutdown, degraded)
• `go test -fuzz=FuzzParseKinetoTrace -fuzztime 30s ./pkg/kineto/` — 506k execs, no failures, 154 corpus entries
• `make lint` — 0 issues
• `make vet` — clean
• `make doc-check` — banned-phrase + section lint clean
• `make generate-fixtures-check` — no fixture drift
• `make generate-check` — `cmd/tracecore/components.go` regenerated, no drift
• `go list -deps ./pkg/kineto/ | grep tracecoreai/tracecore/internal` — empty (no `internal/` leak from `pkg/`)
• `go build ./cmd/tracecore/` — kineto factory linked
• PR D wall-time runs (10-min CPU bench, 100-file soak, 2GB stress) — deferred; code lands, ops run gates and flip non-functional rubrics

Carry-forward

• Non-functional rubric ☑ flips (CPU ≤0.50%, egress ≤0.5Mbps, RSS ≤30MB, peak HeapAlloc <80MB) after PR D wall-time runs.
• `tools/kineto-lint/strace_test.go` (read-only assertion). Trigger: any change to `ingest.go`'s `os.Open` line.
• `pattern_consumer_test.go` compile-time gate for M18 downstream consumer. Trigger: M18 plan starts.
• Upstream `gen_ai.training.*` semconv ratification per NORTHSTARS O4.

🤖 Generated with Claude Code

[trilamsr]

Closing in favor of stacked PRs A→B→C→D per post-shipping review (gap #11). Replacement: m14-a-rfc, m14-b-parser, m14-c-receiver, m14-d-bench. Each reviewable independently; merge order matches stack.