No Tenstorrent equivalent of gpu_clock_state.py: every TT figure on record is clock-unattributed, including #2003

[lu-zero]

.agents/benchmarking.md ("The clock is part of the measurement") makes a
number quotable only with the clock it was taken at, and
tools/bench/gpu_clock_state.py implements that contract for NVIDIA:
sample/stop-summary/judge, min/median/max over retained busy samples,

=30 retained samples, majority-busy fraction >=0.5, within-run spread <=5%,
cross-arm median offset <=1% AND mean offset <=1%, same-boot
(--allow-cross-boot waives exactly the boot id), throttle-reason triage,
refuse-idle-window-with-evidence.

There is NO Tenstorrent sibling, so every TT speed figure ever recorded —
#1604's 10.94-vs-5.34 flip evidence and #2003's inversion (10.822 vs 13.369,
2026-08-26) — is labeled clock-unattributed. The #2003 headline could in
principle be a clock excursion between alternated legs rather than a real
path difference.

Feasibility measured live on thalia/P150 today (tt-smi 6.2.1, pyluwen
0.9.0, tt_umd 0.9.9): tt-smi -s emits a JSON snapshot in 430 ms
including per-device AICLK/ARCCLK/AXICLK, VCORE/TDP/TDC,
ASIC/GDDR temperatures, FAN_RPM, immutable board identity
(BOARD_ID_HIGH/LOW) and firmware bundle versions — 1 Hz external sampling
is directly supportable; importable tt_smi exists for tighter cadence later.

Gaps vs the NVIDIA contract that a v1 must state, not paper over:

• claimed-max clock: not exposed by any tt-smi command => explicit,
  provenance-carrying argument.
• applications/persistence knobs: none exist on TT => NOT APPLICABLE fields.
• throttle reasons: no live bitmap (THM_LIMIT_*/VDD_LIMITS are static);
  context telemetry recorded, throttle-unobservability documented.
• busy/idle discrimination: no utilization counter; proxy = the measured leg
  pid holding an fd on /dev/tenstorrent/*.

First consumer: #2003 — re-adjudicate the inversion with per-arm clock
windows before spending effort on the hybrid-path per-op delta.

Plan: tools/bench/tt_clock_state.py mirroring the helper's CLI verbs,
record schema and thresholds, rule-for-rule; mutation-tested sibling suite;
one wired Qwen3-0.6B leg landing the re-measure. Owner row
BACKEND-TENSTORRENT (spans all TT devices; spec-first).

[lu-zero]

Landed on bench/tt-clock-state (spec → tool+tests → wired-record commits), first consumer re-measured same-day:

• Suite 16/16; live sanity refused an unobserved window exactly as designed.
• Wired leg exposed a real platform fact: the P150 AICLK governor is two-state (800 idle / pegged cap under load). Raw windows therefore refuse within-run spread at 40.74% — kept verbatim, thresholds untouched.
• The busy column (leg-pid holding /dev/tenstorrent fds, recorded live per interval) enables tt_refold_busy.py: all six windows collapse to one distinct value {1350}, spread 0.00%, judge PASS rc=0.
• TT P150 at post-W2c main: the host-hybrid opt-out outperforms the host-free eager DEFAULT by 1.24x — #1604 flip premise inverted #2003 re-adjudicated clock-attributed: default 10.880 vs opt-out 13.645 tok/s median (1.254) at identical cap clocks. Real path difference.

Owed per the spec: verified claimed-max pin, in-process pyluwen sampling, and the spread-scope policy decision.