Mandate 2026-10-20: two-stroke, performance, knock, editing on the engine

.claude/CLAUDE.md

@@ -25,18 +25,27 @@ without the owner's explicit agreement.
    chaotic and neither the behaviour nor the sound is a real two-stroke's. Free
    rein to add what a 2T needs (carburettor, reed and power valves) or to rework
    the ports deeply. Half the chaos was the scavenging model (fixed 2026-10-21)
-   and a carburettor exists (2026-10-22); open are reed and power valves and the
-   3,500/7,000/7,500 rpm holding band, the engine's own and not the dyno's.
-3. **Performance**: test duration, build time, real-time factor.
+   and a carburettor exists (2026-10-22); the 3,500/7,000/7,500 rpm holding
+   band was the 240 Hz frame's alias of the firing ripple, and a power valve
+   exists (2026-10-23). The reed stays a perfect check valve until a measured
+   gap asks for its petals (VISION.md's decision rule).
+3. **Performance**: test duration, build time, real-time factor. JUCE compiles
+   once since 2026-10-23 (Linux CI build 18 -> 4.5 min, Windows 25 -> 32: one
+   serial MSBuild project). Open, and measurable only on the reference machine:
+   the V10-V16 designs below real time (pool spin vs park, duct numerics).
 4. **Knock.** Most engines, even the biggest, survive a boost plus an advance
    that would destroy a real one. Fix the knock PHYSICS, not the damage model.
    The designer's ignition is a template because there is no knock limit to
-   calibrate MBT against, so this item unblocks that one.
-5. **Editing on the engine**: compression ratio by clicking the engine, every
-   turbo visible when two or more are fitted, carburettor parameters by
-   clicking (if one is added), and two exhaust paths joinable into one (a
-   boxer six cannot be made 6-into-1 today). The designer window built on
-   2026-10-19 has never been checked on screen: do that here.
+   calibrate MBT against, so this item unblocks that one. The end gas knocks
+   since 2026-10-22, a stepped control holds the limit, and Douaud-Eyzat's
+   prefactor (x2) puts the catalogue's maps at their limit (2026-10-23). Open:
+   the I5 holds 8-12 deg of retard at its maker's boost; Woschni's combustion
+   term (right for the TDI's BSFC, it stalls lugged engines) was backed out.
+5. **Editing on the engine** -- done 2026-10-23, each checked on screen
+   (journal): compression ratio on the liner/piston, every turbo drawn, the
+   carburettor drawn and jetted by clicking, two exhaust paths joined into one
+   (*Joined* on the ports or a junction), and the designer window. Open: the
+   designer lists valve and injector rows a two-stroke does not use.
 
 **Paused: milestone 1, the sound of the CP2** (MT-07, Arrow exhaust):
 `docs/recording-protocol-cp2.md` holds the method, and the two unverified leads
@@ -125,9 +134,10 @@ masks everything); `EngineLabPhysicsPerfHarness` needs `--idle` beside
 `--trace 1` (`--trace <low rpm>` alone is full-load lugging, not an idle).
 
 **Flags that separate a real effect from an artefact**:
-`EngineLabDynoSweepHarness` `--crank-degrees` and `--coupling-us` resize the
-mechanical substep and the exhaust flush clock, and a gas-dynamic feature stays
-where it is when either changes; `EngineLabGeometrySensitivityHarness` says
+`EngineLabDynoSweepHarness` `--crank-degrees`, `--coupling-us` and
+`--frame-hz` resize the mechanical substep, the exhaust flush clock and the
+240 Hz outer frame, and a gas-dynamic feature stays where it is when any one
+changes; `EngineLabGeometrySensitivityHarness` says
 whether an exhaust change moved the timbre (`shape`) or only the level
 (`level`).
 
@@ -194,7 +204,8 @@ Each measurement is in `docs/journal.md` or the docs archive (tag
 - **Performance**: fork-join per cylinder with woken threads (the resident pool
   runs the pass since 2026-10-20); AVX/AVX2 (dependent-division latency); a
   telemetry cache for listening; state-variable valve-flow reconstruction
-  against clicks.
+  against clicks; whole-project LTO (bit-identical, within noise); `/MP2` on
+  the JUCE project in CI (C1060 beside the app's LTCG link).
 - **Two-stroke**: a 50 mm mesh; coupling the exhaust every substep beyond
   BDC-to-transfer-close, a characteristic valve, finer transfer coupling (none
   of them real time); un-gating exhaust oxidation before scavenging is right
@@ -208,8 +219,8 @@ Each measurement is in `docs/journal.md` or the docs archive (tag
   crankcase); a transit-vs-window scavenging race (monotonic in rpm by
   construction); a scavenging bias from ONE pressure sample at port opening (a
   hunting limit cycle over 4,500-6,250 rpm: integrate the front); "the trapped
-  residual is too low" (a unit error). The 3,500/7,000/7,500 rpm holding band
-  is none of pipe length, settle, step, crank step, flush clock or carburettor.
+  residual is too low" (a unit error). A holding band locked to crank speed is
+  first a frame-rate alias: `DynoSweepHarness --frame-hz` moves it.
 - **Intake**: four inertance formulations (unstable or out of phase); the charge
   bias as the MUSCL limiter, the finite plenum, wall heat or valve linearisation
   (it was the valve nozzle boundary, Courant-dependent); front-loaded or
@@ -225,10 +236,11 @@ Each measurement is in `docs/journal.md` or the docs archive (tag
   fail, 5 more trade the lean first cycle for rich or misfires); seeding the
   high-load correction cell from the idle cell (saves a start at 4 s but stalls
   10 engines of 14 after 25 s of idle) and an asynchronous pulse on tip-in.
-- **Knock, one input at a time** (journal 2026-10-13): the end-gas integral
-  reads the MEAN chamber temperature with a delay 2-5 x slow; isentropic end gas
-  never knocks; plus Douaud-Eyzat it over-knocks (1,000-1,200 K); integral
-  control pulls 12 deg for nothing.
+- **Knock** (journal 2026-10-13, -22): end gas at the MEAN temperature, or at
+  constant cv (1,000-1,200 K); proportional or integral control; Kalghatgi's
+  octane index (boost gains 1-3 octane); an ECU boost cut on knock (I5 363
+  N.m); the I5 at 2.1 bar. A weaker Douaud-Eyzat pressure exponent has no
+  source (pressure matters more in the NTC band).
 - **The other simulator's ingredients, one at a time** (journal 2026-10-13/14):
   output stage, pressure-release outlet, valve-area noise, mild room IR -- none
   kept (blind listening or measurement). Its fine mesh is not shown to matter.

.github/workflows/ci.yml

@@ -22,11 +22,18 @@ jobs:
         run: cmake --preset windows-vs2022
 
       # Two parallel MSBuild nodes: the hosted runner runs out of compiler heap
-      # (C1060) with the default one-node-per-core when JUCE is compiling.
+      # (C1060) with the default one-node-per-core when JUCE is compiling. Two
+      # can still run out (run 98, on run 96's green tree): the build then
+      # resumes on one node, as .claude/CLAUDE.md says to, rebuilding only
+      # what failed.
       - name: Build
         shell: pwsh
         run: |
           cmake --build --preset windows-release -- /m:2 /nologo /verbosity:minimal *>&1 | Tee-Object -FilePath build.log
+          if ($LASTEXITCODE -ne 0 -and (Select-String -Path build.log -Pattern 'C1060' -Quiet)) {
+            Write-Output "::warning::C1060 (compiler out of heap) on two nodes; resuming on one."
+            cmake --build --preset windows-release -- /m:1 /nologo /verbosity:minimal *>&1 | Tee-Object -FilePath build.log
+          }
           if ($LASTEXITCODE -ne 0) {
             Select-String -Path build.log -Pattern 'error [A-Z]+[0-9]+|fatal error' |
               Select-Object -First 30 |

CHANGELOG.md

@@ -52,6 +52,16 @@ file formats, always with an in-memory migration).
 
 ### Fixed
 
+- **The designer's real-time check failed every carburetted two-stroke**
+  ("did not hold 5000 rpm"): it cranked the engine at full throttle, where a
+  carburettor cannot start it. It now cranks with the throttle shut until the
+  engine runs.
+- The designer names the engine from what it is (*Two-stroke Inline Twin*)
+  until a name is typed, instead of keeping *Inline Seven*, and its table's
+  columns are headed *Quantity* and *Value*.
+- A carburetted engine no longer draws, or offers to edit, injectors it does
+  not have.
+
 - A designed V or W engine shares one crank pin between the cylinders of a
   row, as real ones do; each had its own pin. A 90 deg V-twin now fires
   270-450 and a 45 deg one 315-405. Where a shared pin cannot fire six
@@ -74,9 +84,60 @@ file formats, always with an in-memory migration).
   125 cc cylinders a 16.6 mm bore on a single and 33.2 mm on a four. Singles gain
   18-29 % of peak torque; triples, fours and V4s lose the 6-9 % their oversized
   bores gave them; twins are unchanged.
+- **Engines knock.** A boost and an advance that would destroy a real engine
+  now knock: the end gas the flame has not reached is followed as its own
+  zone, compressed from the seal, against an autoignition delay fitted on
+  reference fuels in a knocking engine (Douaud-Eyzat), and the knock's
+  intensity is the pressure that gas adds. The delay's prefactor is
+  calibrated so that the catalogue's own spark maps sit at their knock limit,
+  as a production map does: 10 deg of advance over a map knocks, and a 2JZ
+  run that way at full load with knock control off knocks heavily. The knock
+  control now works as a production one: it pulls the spark back in steps of
+  about a degree while the engine knocks and gives it back at 1 deg/s, so an
+  engine settles at its knock limit instead of sitting in light knock. The
+  Audi I5 gets the intercooler its real engine has.
+- **The compression ratio is edited by clicking the engine**: a cylinder
+  liner or piston offers it beside the bore and stroke, and the running
+  engine takes it (at once or over the ramp).
+- **A carburettor is drawn and tuned in the 3-D view**: on a carburetted
+  engine the throttle body becomes a carburettor (slide tower, float bowl),
+  and clicking it offers its main and pilot jets, needle clip and air screw,
+  taken by the running engine.
+- **Two exhaust paths can be joined into one**: the exhaust ports or a
+  junction offer *Joined*, which gives the other path's cylinders a copy of
+  this path's primary into its collector (a boxer six's two banks become a
+  6-into-1) and restarts the engine.
+- **Every turbocharger is drawn.** With two or more units the 3-D view drew
+  only the first; now each unit an exhaust path drives sits on that path's
+  collector (a flat six's two banks get one each), and clicking one opens its
+  own inspector, titled with its number and its path.
+- **Two-strokes can have a power valve** (`two_stroke.power_valve_*`): it
+  lowers the exhaust port at low speed and lifts as the pipe comes in. The
+  designer fits one to sport two-strokes; a designed 500 cc twin gains 20-39 %
+  of torque between 3,250 and 5,750 rpm and keeps its top end.
+- A two-stroke saved as JSON keeps its scavenging loop length (it came back
+  at the default).
+- A two-stroke twin no longer fails to hold 3,500, 7,000 and 7,500 rpm on the
+  dyno: the brake answered the crankshaft speed sampled at the end of each
+  240 Hz frame, where a firing rate near 240 Hz (or 120) aliases into a slow
+  wander. The dyno now holds the speed averaged over the frame.
+- A designed carburetted two-stroke idles: its pilot jet was sized for about
+  eight times the air a two-stroke idles on, so twins and triples flooded and
+  stalled. The pilot is now sized from the air the idle path passes and
+  follows it as the idle air opens, and each carburettor's idle bypass goes
+  with the cylinder it feeds. Designed 1-4 cylinder two-strokes idle at their
+  1,300 rpm; an RD350-sized design keeps a 35 pilot.
+- A two-stroke's idle no longer hunts: the idle governor runs at half the
+  gain on two-strokes, whose idle air passes the crankcase first. The RD350
+  held 1,175-1,398 rpm with a 5 s period; now 1,284-1,325.
+- Boost no longer burns the charge faster: the flame's turbulence follows the
+  piston, not the intake pressure, so a boosted cylinder burns like a full
+  unboosted one.
 
 ### Changed
 
+- The build compiles JUCE once instead of once per executable: a clean
+  Linux build takes 6 minutes instead of 15 on four cores.
 - Engines with eight cylinders or more simulate their cylinders in parallel,
   with exactly the same result: 22-33 % more real-time headroom (a designed
   6 L V12 at 7,500 rpm 0.71 -> 0.93, the LS3 1.17 -> 1.43, the Merlin

CMakeLists.txt

@@ -131,6 +131,26 @@ if(NOT MSVC)
     enginelab_vendor_headers_are_system(${enginelab_vendor_targets})
 endif()
 
+# JUCE's modules arrive as interface sources, so every target that linked one
+# compiled JUCE again: about 130 of the Linux tree's 390 compile units were
+# JUCE, six to fifteen per executable. They are compiled once here, linked
+# PRIVATE so their sources stay in this library, and their include directories
+# and definitions are re-exported (JUCE's own pattern for a shared code
+# library). Our targets link this, never a `juce::` module directly.
+add_library(EngineLabJuce STATIC)
+target_link_libraries(EngineLabJuce
+    PRIVATE
+        juce::juce_audio_basics juce::juce_audio_devices juce::juce_audio_formats
+        juce::juce_audio_utils juce::juce_dsp juce::juce_gui_extra juce::juce_opengl
+        juce::juce_cryptography
+        juce::juce_recommended_config_flags)
+target_compile_definitions(EngineLabJuce
+    PUBLIC JUCE_WEB_BROWSER=0 JUCE_USE_CURL=0
+    INTERFACE $<TARGET_PROPERTY:EngineLabJuce,COMPILE_DEFINITIONS>)
+target_include_directories(EngineLabJuce SYSTEM
+    INTERFACE $<TARGET_PROPERTY:EngineLabJuce,INCLUDE_DIRECTORIES>)
+set_target_properties(EngineLabJuce PROPERTIES POSITION_INDEPENDENT_CODE TRUE)
+
 add_subdirectory(src/foundation)
 add_subdirectory(src/calibration)
 add_subdirectory(src/events)

cmake/CompilerWarnings.cmake

@@ -1,27 +1,3 @@
-# JUCE's module units are compiled INTO the targets that link them (they arrive
-# as interface sources), so they would inherit the warning set below. They are
-# not our code and they do not build clean under -Wconversion, which drowned
-# EngineLab's own 125 diagnostics in ~400 of JUCE's on the first Linux build.
-# The categories JUCE trips are turned off for those units only; our own
-# sources keep the full set. JUCE's module units are one level deep inside
-# `modules/` by construction, so a single glob finds them.
-function(enginelab_quiet_vendor_units)
-    if(MSVC OR NOT juce_SOURCE_DIR)
-        return()
-    endif()
-    get_property(units GLOBAL PROPERTY ENGINELAB_VENDOR_UNITS)
-    if(NOT units)
-        file(GLOB units "${juce_SOURCE_DIR}/modules/*/*.cpp")
-        set_property(GLOBAL PROPERTY ENGINELAB_VENDOR_UNITS "${units}")
-    endif()
-    if(units)
-        # A source property resolves in the directory that defines the target,
-        # which is this function's caller: a function is not a new scope.
-        set_source_files_properties(${units} PROPERTIES COMPILE_OPTIONS
-            "-Wno-conversion;-Wno-float-conversion;-Wno-missing-field-initializers;-Wno-shadow;-Wno-comment")
-    endif()
-endfunction()
-
 # A vendored *header* is compiled as part of whatever includes it, so no source
 # property can reach it. One line of JUCE's `juce_UMPUtils.h` broke fourteen of
 # our translation units on the first Linux build with warnings as errors, and
@@ -66,6 +42,5 @@ function(enginelab_set_warnings target)
         if(ENGINELAB_ENABLE_WARNINGS_AS_ERRORS)
             target_compile_options(${target} PRIVATE -Werror)
         endif()
-        enginelab_quiet_vendor_units()
     endif()
 endfunction()