[tannevaled]

Four units.

PR	area	gain
#645	the line map	language/line_spec 0/6 → 6/6, real lines in backtraces, MRI-shaped frame labels
#640	Kernel#sleep + the fiber scheduler, File.open's mode/flags/perm contract, IO::Buffer.map	+90 across core/file, core/io, core/kernel; sleep_spec 0 → 15
#639	regexp search anchoring and encoding, Refinement#import_methods, pack/unpack
go-ruby-parser v0.3.0	ast.Program.Lines	the prerequisite for #645

The wall was one repository down

__LINE__ evaluated to 0 everywhere, and 101 examples across five directories were blocked on it — const_source_location, every Thread::Backtrace::Location spec, each_caller_location, warn(uplevel:), line_spec. Three earlier agents each reported it as the thing they could not get past, and each assumed the fix was in the compiler.

It was not. The parser's AST carried no position information at all: 55 node types, no line or column field anywhere, and the lexer — which knows every token's line — discarded it at the AST boundary. No amount of compiler work could have produced a line. That is why this took a parser release (ast.Program.Lines, v0.3.0) before the VM half could start.

The table itself is MRI's shape, not a line per instruction: insns_info[] beside positions[], written in emit() so a statement emitting nothing contributes no entry and one emitting twenty contributes one (iseq.c v3_4_0:673, read back by get_insn_info_binary_search at :2160).

Measurement

scripts/conformance/rubyspec/run.sh, pinned corpus 87b1631, 16-way, TIMEOUT=25, three consecutive runs:

passing examples: 22153   (files loaded: 2192, failed to load: 14)
passing examples: 22152
passing examples: 22152
frozen floor:     21970

Floor set twelve below the lowest run — the margin rule the last three bumps used: enough for the run-to-run jitter, not enough to hide the loss of a whole spec file. #615 is still open and still costs a file when it fires.

Worth recording from this wave

• A regeneration of the frozen prelude came back byte-identical, which looked like good news and was not: the AOT emitter was dropping Lines, so every AOT-compiled frame would have reported line 0 — the old behaviour reintroduced silently on that one path. TestEmbeddedPreludeMatchesSource caught it.
• A benchmark improvement of −0.2% turned out to be noise on a benchmark whose own spread is 870–1037 ms; fifteen interleaved runs put the real figure at +7.5%. The code comment that had recorded −0.2% was corrected, so nobody re-derives it.
• IO.select carried a NaN guard that could never run — the coercion it calls raises on NaN itself rather than returning one. It read as an uncovered branch; it was dead code, and is deleted.

🤖 Generated with Claude Code