[Rovey]

Refs #84. The implementation of the design in #86, for the Undervolt profile. The design was approved by the owner, including two deviations listed below. Hardware check 111 passed on the reference RTX 4070 (2026-11-14); check 112, a card that is hot at stock, has not run.

What changes

Core (src/core/, pure logic, unit-tested):

• run_stability also averages the last half (kCountedShare = 0.5) of a measurement that ran its full length, in StabilityResult::counted: score, power, core and memory clock, temperature and fan speed. A measurement that ended early has no counted part. counted_line shows a measurement as it was compared: the averages, then the whole run's peak and last fan reading, each named.

• judge_against_stock (core/undervolt.hpp) decides whether the Undervolt profile keeps its result. It needs a valid comparison: both sides ran their full length under the same fan control. Then:

  • not slower: at least kNotSlowerScore (99 %) of the stock score;
  • no thermal regression: not more than kKeepCoolerC (2 C) warmer unless at least kKeepQuieterPts (5 points) quieter, and not more than 5 points louder unless at least 2 C cooler;
  • at least one benefit: power at most kKeepPowerShare (97 %) of stock's, or 2 C cooler, or 5 points quieter.

  Higher power alone never discards a result, and finishing the soak alone never keeps one. The tolerances are provisional named constants.

App (src/app/):

• measure_stock can drive the fans with a curve (FanDriver, handed back by FanRelease).
• run_undervolt, when not on top of a saved tune:
  • measures stock on the curve the run uses on a card that stays under the profile's limit;
  • when that peak raises the ceiling and so moves the curve up, measures stock once more on the moved curve, and compares the result with that second measurement (the first still sets the ceiling and the clock to keep);
  • stops before the search with not compared: the measurement at stock stopped at N C after S s when the measurement to compare with did not run its full length;
  • after the soak, judges the result. One that is not kept is put back to stock and not saved. RESULT: not applied -- <reason> gives the reason with both sides' numbers.
• A kept result shows and saves exactly what it was judged on:
  • the result lines: averages of the last 150 s: score= core= power= temp= fan=; whole 300 s: peak= fan at the end=;
  • the result screen: a row for each average (speed, core clock, power, temperature, fan speed), plus Peak temperature, whole run and Fan speed at the end apart. The headers read Before (stock, fans on the app's curve) and After (fans on the app's curve), with a note under the table;
  • the saved against_stock: MeasuredRun gains counted_s, avg_temp_c and avg_fan_pct. They are written only for such a comparison, so every other profile file is written as before; a file without them reads as before. The Dashboard shows Temperature, average and Fans, average next to Peak temperature.
• The Optimize page adds, for Undervolt: Up to 5 minutes longer when the card runs hotter than 75 °C at stock: it is then measured at stock a second time, on the fan curve the run uses.

Docs:

• README (the Undervolt rules and the time estimate);
• a design-decision row in docs/development.md;
• hardware checks 111 and 112 in docs/hardware-checks.md, marked Not run yet.

Review fix (553ff50)

The verdict used the counted half's average temperature and fan speed, but the result lines, the result screen and the saved comparison showed the whole run's peak temperature and last fan reading in their place (counted_view copied only score, clocks and power). counted_view is removed, and every place now shows and keeps the averages the verdict used, with the peak and the last reading named apart (see above). The thresholds and the comparison are unchanged.

Deviations from the design, approved by the owner

• The speed profiles and All in one keep their measurement at stock on the driver's fans. Changing it would change the "before" their own slower-than-stock rule and the All in one table compare with; that is a change of its own.
• An undervolt on top of a saved tune (gao --undervolt --on-overclock, the undervolt step of All in one) is not judged against stock. Its clock is lower than the overclock's on purpose, so "not slower than the reference" would discard it every time.

Tests

• cmake --workflow --preset ci (Release, /W4 /WX): no warnings.
• core_tests: 514 test cases, 11794 assertions, all passed (was 501). New tests cover:
  • the counted half: the averages, an early end, a card without a fan reading;
  • counted_line;
  • the new against_stock fields: written only with averages, read back, and ignored unless all three are there;
  • judge_against_stock with the RTX 4070 and GTX 1080 Ti numbers, every rule at its boundary, a missing reading, other fan control, and a soak without a counted part.
• app_tests: 55 test cases, 320 assertions, all passed (was 52). New tests cover:
  • the result lines of a judged undervolt: averages, not the whole run's numbers, with the peak apart;
  • the result lines of a discarded one;
  • the saved comparison (counted_against_stock) keeping the judged averages through a profile-file round-trip.

Hardware check 111: passed on the reference RTX 4070 (2026-11-14)

Build f2b235e, driver 617.42, fan curve 50 C 52 %, 70 C 52 %, 75 C 100 %.

Run	Applied	Stock (last 150 s): score / power / temp / fan	Result (last 150 s)	Against stock
1	990 mV, kept	5727.0 it/s / 198 W / 70 C / 62 %	5702.1 / 159 / 67 / 52	-0.4 %, -19.7 %, -3 C, -10 pts
2	990 mV, kept	5728.2 / 198 / 70 / 62	5701.3 / 158 / 66 / 51	-0.5 %, -20.2 %, -4 C, -11 pts
3	990 mV, kept	5725.7 / 198 / 70 / 61	5701.0 / 159 / 67 / 52	-0.4 %, -19.7 %, -3 C, -9 pts
window	990 mV, kept	5710 / 198 / 70 / 61	5682 / 158 / 67 / 52	-0.5 %, -20.2 %, -3 C, -9 pts

• No second measurement at stock, as expected: the card peaked at 71 to 72 C on the curve.

• The scatter, twice the range over the three command-line runs (one tune, so it is measurement noise):

  	Score	Power	Temperature	Fan
  Stock	0.09 %	0 %	0 C	2 pts
  Result	0.04 %	1.3 %	2 C	2 pts
  Tolerance	1 %	3 %	2 C	5 pts

  Every tolerance holds. The values stay as they are. Power, temperature and fan are whole numbers, so a range of 1 is the smallest step; the temperature tolerance sits at twice that step.

• The window and the Dashboard showed the judged averages with the right headers, rows and notes. The window run scored about 0.3 % lower on both sides; it is left out of the scatter.

• Not seen on hardware: a result that is discarded, and not compared:. These are covered by unit tests only.

• Found during the check, not caused by this PR: a kept result also installs the build under test as the copy the logon task starts (update_logon_copy). This is existing behaviour. The procedure in AGENTS.md and checks 111 and 112 now says to put the installed release back afterwards; that was done here.

Hardware check 112: not run

It needs a tester with a GTX 1080 Ti or RTX 3080 Ti that runs hot at stock. A local test build, version 0.4.10-check112, has been prepared with instructions. It is not published.

Not to be merged before the owner's review.

🤖 Generated with Claude Code