igneum-labs d55dfd003f 0.3.25: the AMD core-clock and power-cap knob for Ember Tune (main's order of 8 October 2026: PC 1's RX 9070 XT measured 18.9 MH/s at 202 W and stopped, "no knob on AMD"; the card is latency-bound on the whole class v4 ladder, so a core-clock cap is the lever as on NVIDIA). Two findings behind the stop: the app's AMD lever (--tune, --set-gmax, --set-plimit, --reset in igneum-gpu-telemetry: ADLX IADLXManualGraphicsTuning2 and IADLXManualPowerTuning on Windows, pp_od_clk_voltage and hwmon power1_cap on Linux) was built on 5 October (620296b2) and never left branch opencl-rdna4-telemetry, so the kit's exe answers no tune line and every AMD tune fell to "measure only"; and the 9070 XT's max clock is an OFFSET range (gmax 0, gmax_range -500 1000), which the engine read as "no clock knob". Now: proto-opencl/gpu-telemetry.c takes 620296b2's tool whole (a superset of the shipped one); src/ember.rs amd_knob turns the tune line and the card's stock clock under load into the knob (clock ladder stock down to stock + gmax_min, power ladder on the percent scale 100 + plimit range; absolute-MHz drivers pass through), amd_limits, amd_gmax_arg (the apply sends the offset), and "not available (<reason>)" with nothing set for no AMD device, a tune line that read an error, Linux (root under /sys, a later cut), or a stock clock not yet known; Plan::full takes a narrow range (the floor above the 45 percent rung) as the fine ladder alone in 100 MHz steps, so the 9070 XT runs 2870, 2770, 2670, 2570, 2470 under a 2,970 MHz stock, the stop rule at the knee or a faulted row, lock_result and the lock_* fields as on NVIDIA. ADLX manual tuning needs no elevation (PC 1, 5 October 2026, an unelevated job), so the no-prompt rule holds with no Power Helper verb; nothing of the engine runs elevated. engine.rs carries the whole tune line in TuneProbe.amd_tune and the card's amd_stock_mhz and amd_gmax_offset. Tests known-failed first: the_amd_knob_reports_not_available_with_its_reason_and_sets_nothing, the_9070_xt_gets_a_power_ladder_and_a_clock_ladder_from_its_stock_clock, the_9070_xt_ladder_locks_at_the_knee (a declared ladder in the card's shape, not a measurement). Owed: the kit's igneum-gpu-telemetry.exe rebuilt from this source (MSVC on a PC or build-1, the ADLX SDK at vendor/adlx beside the tree), then the first measured grid on PC 1 by job when the hash lane's queue is clear ... Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>		2026-10-08 08:42:15 +00:00
..
igneum-app	0.3.25: the AMD core-clock and power-cap knob for Ember Tune (main's order of 8 October 2026: PC 1's RX 9070 XT measured 18.9 MH/s at 202 W and stopped, "no knob on AMD"; the card is latency-bound on the whole class v4 ladder, so a core-clock cap is the lever as on NVIDIA). Two findings behind the stop: the app's AMD lever (--tune, --set-gmax, --set-plimit, --reset in igneum-gpu-telemetry: ADLX IADLXManualGraphicsTuning2 and IADLXManualPowerTuning on Windows, pp_od_clk_voltage and hwmon power1_cap on Linux) was built on 5 October (620296b2) and never left branch opencl-rdna4-telemetry, so the kit's exe answers no tune line and every AMD tune fell to "measure only"; and the 9070 XT's max clock is an OFFSET range (gmax 0, gmax_range -500 1000), which the engine read as "no clock knob". Now: proto-opencl/gpu-telemetry.c takes 620296b2's tool whole (a superset of the shipped one); src/ember.rs amd_knob turns the tune line and the card's stock clock under load into the knob (clock ladder stock down to stock + gmax_min, power ladder on the percent scale 100 + plimit range; absolute-MHz drivers pass through), amd_limits, amd_gmax_arg (the apply sends the offset), and "not available (<reason>)" with nothing set for no AMD device, a tune line that read an error, Linux (root under /sys, a later cut), or a stock clock not yet known; Plan::full takes a narrow range (the floor above the 45 percent rung) as the fine ladder alone in 100 MHz steps, so the 9070 XT runs 2870, 2770, 2670, 2570, 2470 under a 2,970 MHz stock, the stop rule at the knee or a faulted row, lock_result and the lock_* fields as on NVIDIA. ADLX manual tuning needs no elevation (PC 1, 5 October 2026, an unelevated job), so the no-prompt rule holds with no Power Helper verb; nothing of the engine runs elevated. engine.rs carries the whole tune line in TuneProbe.amd_tune and the card's amd_stock_mhz and amd_gmax_offset. Tests known-failed first: the_amd_knob_reports_not_available_with_its_reason_and_sets_nothing, the_9070_xt_gets_a_power_ladder_and_a_clock_ladder_from_its_stock_clock, the_9070_xt_ladder_locks_at_the_knee (a declared ladder in the card's shape, not a measurement). Owed: the kit's igneum-gpu-telemetry.exe rebuilt from this source (MSVC on a PC or build-1, the ADLX SDK at vendor/adlx beside the tree), then the first measured grid on PC 1 by job when the hash lane's queue is clear	2026-10-08 08:42:15 +00:00
mac	The opening window size, one rule for both hosts (window-22, 7 October 2026, the project lead on PC 2: "why did the app open super wide?"): scaling-21's 80 percent of the work area scaled above 1920 wide opened 2752 by 1152 on a 3440-wide area and 3072 by 864 on a two-monitor span, and a remembered WindowW/WindowH could carry that into later starts; now app/windows/opening_size.h: the PRIMARY monitor's work area (MonitorFromPoint primary, never the virtual desktop), width 80 percent capped at 1440 up to a 1920-wide display and 1920 beyond, height from the width at 16:10 bounded by the work area, never wider than 1.6 times the height, never under 900 by 600, never over the work area; a remembered frame kept only when it fits the work area, the minimum and the aspect cap, and never saved when it would not; the Mac window mirrors the rule on NSScreen.screens.first and discards an autosave frame that fails it; opening_size_test.c compiles on the gate with cc (known-failed first: it could not compile before the header, and 0.3.21's rule fails the 3440 by 1440 and the 3840 by 1080 cases); host.cpp touched (the shipper's named line)	2026-10-07 18:34:47 +00:00
windows	0.3.24: version strings first (rule 15; the six places), the tree after 0.3.23's cut: class v5 by height on Devnet 3, the chain id at the v5 floor, the dashboard width caps, the pool daemon's reconnect and counters	2026-10-07 21:37:44 +00:00