Scan only a command whose invocations are all compiles

[typeless]

Summary

Two commits, split because two files carry both changes:

1. Split a scan command into invocations before reading it — a scan is built
   by walking a command's words, but two of the four walks that do it decided
   what a word meant from its position in the whole list rather than from the
   invocation it belongs to: both matchers searched for the compile flag across
   every word, and both builders folded source words in from wherever they
   appeared. This splits the command into invocations in dep_words.cpp,
   beside the separator notion the scanners already share, and re-expresses
   all four call sites over it. Along the way, is_command_separator is
   renamed to is_flag_barrier — it never meant "separates commands", it
   means "the scan reads no flags past this word", which is why redirections
   belong in it and why it is the wrong set to divide invocations on. No
   behaviour change.

2. Scan only a command whose invocations are all compiles — a scan runs
   from the rule's directory with the rest of the command stripped, so a word
   taken from a later invocation is read in a state that invocation never had.
   A chained command like gcc -c a.c -o a.o && cd sub && gcc -c b.c -o b.o
   could emit a scan that reads a same-named file in the wrong directory,
   producing headers stale against the rule and never rebuilding on the one it
   actually reads. The fold wasn't limited to compiles either — a trailing
   && rm junk.c contributed junk.c, and a later invocation's -c could turn
   a link into a false compile match. Now a scan may carry a word only from an
   invocation putup can reproduce, and it can reproduce an invocation only
   when it and every invocation before it is a compile it recognizes.

RED was quoted first: seven tests, plus one earned by pair-partner review
during the closing pass. The whole suite passes in one process: 434 ~[e2e]
cases plus 395 [e2e] cases. A Sonnet pair-partner review caught a real
blocker — a regression on trailing separators — which was fixed before this
branch was considered done.

Closes #356

Test plan

• RED quoted before production code changed
• Whole suite passes in one process (~[e2e] + [e2e])
• make format-check, make spec-check clean
• Pair-partner (Sonnet) review, one blocker regression on trailing
  separators caught and fixed
• Measured cost in tree: none — no example rule is compiler-first with a
  separator

[github-actions]

PR metrics

Performance (gcc example, Linux)

Workload	Instructions	CPU time	Page faults	D1 miss	LL miss	Wall	Peak RSS
parse	1768 M (+0.3%)	0.65 s	14.7 k (+1.4%)	0.7%	0%	0.65 s	33.2 MB (+1.0MB)
dry-run	2210 M (+0.3%)	0.7 s	15.6 k (+0.6%)	0.8%	0%	0.724 s	36.5 MB (+0.4MB)

Deterministic signals: instructions (cachegrind-simulated instruction reads — exact across runs, no PMU needed), page faults, peak RSS, and the cachegrind D1/LL miss rates. CPU time is user+sys from time(1).

Internal statistics (gcc example, up-to-date dry run)

Metric	Value
Tupfiles parsed	24
Commands	3545
Commands scheduled	0
Files checked	5834
Files changed	0
Files in index	6102
Graph edges	381415
Index size (bytes)	7683630
Implicit deps	344126
Hash computations	0
Hashes skipped (stat cache)	5833
Stat calls	5885
Parse time (ms)	564.7 (-0.6%)
Total time (ms)	698.8 (-0.6%)
Runner CPU	AMD EPYC 7763 64-Core Processor

Counters from putup -n --stat on the fully-built gcc example (up-to-date dry run): deterministic work measures — a jump in commands scheduled, hash computations, or stat calls is a real behavior change, not noise. Timings are the minimum over repeated runs, compared only against a baseline from the same CPU model; the counters are the regression signal.

Binary size (Linux)

Binary	.text	.data	.bss	File
putup	564 KB (+0.3%)	2.3 KB	98.7 KB	670.9 KB (+0.7%)

Code churn (whole codebase, last 30d)

Files	Lines written	Still present	Churned	Churn rate
140	11234	9575	1659	14.8%

Of the lines written across the codebase in the last 30 days, how many are already gone — work that was written and then discarded or rewritten inside the same window. This is the state of the tree including this PR, not a measure of the PR itself. Only code we write is counted: tests, examples, vendored and generated files, CI plumbing and prose are excluded. 5065 lines were deleted in the window in total, most of them older than it.

Where the churn is

File	Lines written then discarded
src/cli/cmd_build.cpp	653
src/graph/scanners/clang_cl.cpp	155
src/cli/context.cpp	82
src/graph/builder.cpp	72
src/core/layout.cpp	69
src/core/new_delete.cpp	63
src/graph/scanners/dep_words.cpp	63
src/platform/file_io-posix.cpp	42
include/pup/index/format.hpp	35
src/cli/cmd_clean.cpp	35

Test coverage (lines)

Overall	Median file	Min file	Max file
88%	96.2% (-0.2pp)	14.7% include/pup/parser/token.hpp	100.0% include/pup/core/arena.hpp

102 files · 16512/18758 lines covered

Deltas vs main@bb2c7877f.

Updated for 12bfb89