feat(profiling): triage --axis compile — classify compile drift, close the arc

[Jammy2211]

Overview

Phase 3 of 3 of the compile-axis arc (phase 1 #218/#219; phase 2 autolens_profiling#103,
PR #220 + autolens_profiling#104).

Phases 1–2 make compile drift visible; this makes it actionable, answering the
question that decides what anyone does next: who owns this?

Plan

• Add --axis compile to triage, classifying each ingest finding.
• Compare a suspect warm row against its own cold scale — the yardstick that makes
  cache-regression a measurement rather than a guess.
• Route library findings to bug/ via intake; never debug the library in the profiling
  repo (the agent's existing boundary, unchanged).
• Close the arc in AGENTS.md: compile-time profiling moves out of Future modes,
  and Boundaries records that release-validation script cost stayed with hygiene.

Classifications

classification	signal	actionable
cache-regression	warm compile returned to its own cold scale	yes — config/stack, never the library
autotune-regression	GPU compile up ≥10× with no cold-scale match	yes — check the flag reaches XLA
library-regression	growth on an unchanged key, no other explanation	yes — bug/ via intake
host-load	measuring host's 1m load average was high	no — re-measure idle first
expected-recompile	key differs from a pin only by jax_version	no — one recompile is by design
new-machine / new-precision / new-cell	key is simply unpinned	no — pin it

Acceptance

• Every ingest finding classified into exactly one outcome, with evidence.
• A synthetic warm-reverting-to-cold row classifies as cache-regression —
  verified against a copy of the real workspace, so the cold reference is real
  data: a warm vag row moved 1.622 s → its own 34.592 s cold cost and classified
  correctly with that evidence quoted.
• A jax_version bump classifies as expected-recompile, not a regression.
• Nothing written to autolens_profiling (asserted by mtime comparison).
• AGENTS.md Modes + Boundaries updated.

Note on the prompt's original table

Two of the five categories the prompt specified — drift from a jax_version bump, and
drift from changed measurement conditions — cannot reach triage as drift by
construction, because both are different comparability keys and ingest reports them
as unpinned rather than drifted. That's the design working rather than a gap. They're
still classified so nothing vanishes, but they're never regressions. host-load was
added in their place, which the prompt didn't list but the data supports — it's what
host_state was added for in phase 2.