The sysmon core needs the multiline map (line -> first line of its
multi-line statement) to resolve branch events.  It got it by running a
full PythonParser.parse_source() per file — an ast.parse, a full
tokenization, a compile() via ByteParser, and AST walks — behind a
module-level functools.lru_cache(maxsize=20).  Test suites tracing more
than 20 files thrash that cache and re-parse files repeatedly (the
pyca/cryptography suite traces 192 files; see issue coveragepy#2172 for a report
of sysmon branch mode being slow on such suites).

Three changes:

- Make multiline_map_from_tokens() in parser.py the one place the map
  is computed: PythonParser._raw_parse tokenizes once into a list, gets
  the map from the shared builder, and its own token loop keeps only
  the exclusion and indent bookkeeping (its in-flight first_line
  tracking stays, since the exclusion logic needs the statement start
  before the map entry for the current statement exists).

- Use multiline_map_from_text() in the sysmon core instead of a full
  parse.  This is ~4x cheaper (0.34s vs 1.48s for the 192 files of the
  cryptography suite) and produces identical maps (verified on 445
  files: coverage's own source and tests, the stdlib, cryptography and
  its tests).

- Cache the maps in a plain dict on the SysMonitor instance, unbounded,
  so each traced file is tokenized at most once per run.  The cache
  dies with the tracer, which also removes the cross-run staleness a
  module-level cache can have.

Measured on the cryptography suite (Python 3.14.2, branch mode, wall
time best of 3, base 35.44s): 47.90s (+35.2%) before, 46.07s (+30.0%)
after, with byte-identical coverage data.  The parse/report phase pays
one extra pure-Python pass over the already-materialized token list
(parse_source() over the 192 files: 1.75s -> 1.95s, ~+1ms per file),
which the reporting-phase caches already amortize.  Most of the
remaining measurement overhead is the branch_trails() analysis,
addressed separately by the sysmon-lazy-branch-resolver branch; the
two changes compose to ~0-2%.

Verified: parser results (statements, excluded, raw_excluded,
multiline_map) identical to released main on 445 files with a realistic
exclusion regex; coverage's own test suite failure set identical to
released main.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0137DLbSXfm5v5bhCcz7xEKM

Add directed tests of multiline_map_from_text() for each statement shape, plus a parametrized check that it always matches the map PythonParser produces, since branch arcs are attributed with one and reports are keyed by the other.  Also test compute_multiline_map()'s fallbacks for missing, non-Python, and badly indented files, and that SysMonitor computes each file's map at most once per tracer instance.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0137DLbSXfm5v5bhCcz7xEKM