Run the compiler stages in a browser, and build the constants list ourselves

[tamnd]

Closes #77. Closes #78.

Both issues came out of the Pyodide measurement in #76, and they are the same bug seen from two sides, so they are fixed together.

The bug

compiler_codegen collects the constants it comes across and hands them back in its metadata. The WebAssembly build does not: the dictionary has argcount, kwonlyargcount and posonlyargcount and nothing else. pyxray.compiler.stages passed metadata["consts"] straight into optimize_cfg, so that line raised KeyError in a browser and the middle compiler stage did not run there at all. T05 and the pipeline widget both go through that function, and both are meant to run in a tab with nothing installed.

The fix

stages builds the constants list itself now, from the instruction sequence it is about to pass in, and records on Stages.constants_known whether the values were the real ones. The optimizer runs everywhere. On a build that hands the values over, nothing changes at all.

The part that needs saying out loud

The optimizer without the values is not the optimizer doing nothing. It cannot fold 6 * 7 into 42, and on something like while 0: it makes a different decision than it would with them, so the right hand column in a browser is a real answer to a question the reader did not ask. Three places say so now: what_the_optimizer_did adds a line under the two columns, the pipeline widget marks that pane and explains it in a note, and T05 prints compiler.constants_available() next to compiler.available() and has a paragraph about what a False there means. The last pane and the last cell are the finished code object, which comes from the ordinary compile() and is the real answer on every build.

The placeholder is a small class rather than None, which matters more than it looks. The optimizer rewrites a LOAD_CONST of a None into the shorter LOAD_COMMON_CONSTANT form, so padding with None would show a browser reader LOAD_COMMON_CONSTANT where their source has a 6, which reads as an optimization and is not one. A placeholder it cannot recognise leaves LOAD_CONST 0 sitting there, which is honest.

The safety half, which is #78

optimize_cfg reads that list by index. Hand it one shorter than the largest index in the sequence and a native interpreter raises a tidy ValueError, while the WebAssembly build reads past the end of its memory and does not come back. There is no exception, so there is nothing to catch, and in a notebook it takes the kernel and whatever the reader had done. That is measured, in probes/pyodide/pyodide.json under optimize_cfg_short_consts.

So the list is built inside pyxray.compiler every time and taken from nobody. The pipeline widget, which is the one place a reader types their own source and therefore the easiest way to reach this, has no other route to the optimizer. There is a parametrized test that drives seven real sources through the widget's own path, watches the call, and checks the list was long enough, because the property that has to hold is the length rather than the absence of a crash on an interpreter where a short list would only raise. The rule is written down at the top of pipeline.py so the next person to add a parameter there knows why they should not.

wasmprobe

The optimize_cfg check was described as "the way pyxray calls it", and pyxray does not call it that way any more. It drops from tier0 to info and loses its accepted excuse, and its cost line now says what it used to cost and why it is still worth asking every run. optimize_cfg_direct takes over the tier 0 wording and builds its list the way pyxray does, off dis.hasconst rather than a hard coded LOAD_CONST. The recordings and the report are rebuilt.

No check carries an accepted gap now, which is the good outcome and not a reason to delete the mechanism, so that test builds a check with one rather than asserting the real list is not empty. probes/pyodide/decision.md and the browser section of CONTRIBUTING.md are updated to match.

Checks

just check is green: lint clean, 1428 passed and 3 skipped, 535 citations, 12 lessons and 12 notebooks up to date and running, the probe still reporting 15 checks with nothing blocking, and 5 animations.