feat: --unwind auto lazy CFI compile

[dpsoft]

Summary

Implements lazy CFI compile for --unwind auto -a. Replaces today's eager AttachAllProcesses (compiles every binary visible at startup, ~30 s on a typical desktop) with a lazy variant that populates pid_mappings only at startup and defers per-binary compile until a sample fires inside an unprepared binary.

Validated bench results on this hardware:

Metric	Eager (--unwind dwarf)	Lazy (--unwind auto)	Δ%
p50 wall time	34,618 ms	8,141 ms	−76.5%
p95 / max	69,683 ms	40,982 ms	−41.2%
Warm-state runs (2-5)	34-37 s	6-8 s	78-83%

5× faster from the second invocation onward — the operator-attaching-to-running-host case the doc was concerned about.

Architecture

• BPF (bpf/unwind_common.h): new cfi_miss_events ringbuf (64 KB) + cfi_miss_ratelimit LRU_HASH (4096 entries, 1-sec rate limit per (pid, table_id)). Walker probes cfi_classification_lengths[table_id] before classify_rel_pc; if missing, emits a miss event and falls through to FP path. The existing FP_LESS+cfi_lookup miss path also emits as a fallback for the rare case of compiled-but-incomplete CFI.
• Userspace drainer (unwind/dwarfagent/miss_drainer.go): consumeCFIMisses goroutine reads ringbuf, dedupes per-(pid, table_id), poisons after 3 failures, calls tracker.AttachCompileOnly(pid, path) to compile on demand. MissStats exposes counters.
• Tracker (unwind/ehmaps/tracker.go): new EnrollWithoutCompile (populates pid_mappings, no refcount) + AttachCompileOnly (compiles + populates cfi_*, takes refcount, no pid_mappings write).
• ScanAndEnroll (unwind/ehmaps/scan_enroll.go): walks /proc/* and enrolls every binary with a shared build-id cache. Cached so 1,000 unique binaries get ~1,000 build-id reads, not ~270,000.
• Mode dispatch (unwind/dwarfagent/agent.go): new Mode enum + NewProfilerWithMode. Per-PID forces ModeEager (compile cost is 1.4% there, lazy buys nothing). Existing NewProfiler/NewProfilerWithHooks delegate with ModeEager — backward compatible.
• --unwind dispatch (perfagent/agent.go): auto → ModeLazy; dwarf → ModeEager (escape hatch). Off-CPU stays eager (v1 non-goal per spec).
• Bench harness (bench/cmd/scenario/main.go): --unwind {auto|dwarf} flag for before/after diffs. schema.Config.UnwindMode records the chosen mode.

What's preserved

• --unwind dwarf retains today's eager behavior. Same code path, plus one map lookup per frame in walk_step (~tens of ns).
• --unwind fp unchanged.
• --pid N unchanged regardless of --unwind value (compile cost is already negligible).
• All existing unit + integration tests pass; off-CPU profiler unchanged.

What was caught and fixed during execution

• Spec reviewer caught a legacy wg.Add(1) + go func() pattern (commit 5034cfe0) — converted to Go 1.25's wg.Go(...).
• Integration test failed on first run with MissStats.Received == 0 — exposed a real spec bug. Original spec said the BPF emit point was "FP_LESS + cfi_lookup miss"; in lazy mode cfi_classification isn't populated so classify_rel_pc returned MODE_FP_SAFE and the FP_LESS branch was unreachable. Added a cfi_classification_lengths probe before classify_rel_pc (commit 8b9dba69). Test passed on second run with Received=32, Resolved=32.
• Code reviewer caught a LastLatencyNs clock mismatch (time.Now() vs bpf_ktime_get_ns()); fixed by using unix.ClockGettime(unix.CLOCK_MONOTONIC, ...) (commit 2f13d419).
• Off-CPU scope creep in Task 9 (added NewOffCPUProfilerWithMode, would have nil-panicked) — reverted (commit 3e9af28e).

Test plan

• make test-unit passes (7 packages, 0 failures).
• BenchmarkScanAndEnroll_BuildIDCacheHit reports buildid_reads/op = 5.000 for 100 PIDs × 5 binaries — proves the cache caps reads at K, not N×K.
• TestLazyMode_FiresAndCompilesOnMiss (caps-gated) PASS: Received=32, Resolved=32, PoisonedKeys=0 after 5-second window.
• make bench-scenarios system-wide-mixed --processes 30 --runs 5 for both --unwind dwarf and --unwind auto. Diff shows −76.5% p50 reduction.
• Soak test on a high-fork-rate workload to confirm rate-limit + LRU eviction behaves under load (follow-up if needed).

Companion docs

• Spec: docs/superpowers/specs/2026-04-27-unwind-auto-lazy-a2-design.md (with the validated-results section).
• Plan: docs/superpowers/plans/2026-04-27-unwind-auto-lazy-a2.md.
• docs/unwind-auto-refinement-design.md updated to mark A2 as implemented.

Follow-ups (deliberately deferred)

• Off-CPU lazy mode (v1 non-goal — small ringbuf cost on off-CPU sessions documented inline).
• Userspace-side compile parallelism (current drainer is serial; bench shows it's fast enough).
• MissStats.LastLatencyNs exposure to a histogram / external metric.