feat: expose whether the version profile is exact

[CameronBrooks11]

Closes #61.

get_profile() already returned (profile, exact). OpenRocketInstance.__init__
consumed exact, logged a warning, and discarded it — so no caller could tell an
exact profile from a nearest-older fallback. SimulationPool had the same defect
one layer up: it builds an instance to validate the jar and fire that warning once
in the parent, then threw the instance away.

A log line is not a surface. An application that raises its log level silences the
only signal, leaving no exception, no return value and no attribute. And the obvious
reconstruction is wrong: parse_version maps 24.12.RC.01 to (24, 12), an exact
match whose version strings differ, so or_version != profile.version_string reports
a fallback on every RC and point release.

Adds

• OpenRocketInstance.profile_exact — True when a profile for the jar's exact
  version is checked in, False on a nearest-older fallback.
• SimulationPool.profile_exact — the same verdict on the pool, which is the
  monte-carlo path where running on stale enums matters most.

Both are additive. tests/test_api_surface.py freezes module-level __all__, which
is unaffected.

Docs

Four pages described the fallback as a warning, which would have left the fix
undiscoverable to anyone not reading source. docs/guides/multi-version.md — the
page a caller reads for this question — now shows the flag with a worked example and
states why the string comparison is not equivalent. docs/index.md, README.md and
docs/maintainers.md updated to match.

The class docstring uses :ivar: rather than an RST definition list: mkdocs.yml
sets docstring_style: sphinx, and the definition-list form was passed through as
description text and collapsed by Markdown into a single run-on paragraph.
mkdocs build --strict does not catch that.

Verification

5 new test functions, 7 cases with parametrisation. Reverting both assignments and
keeping the tests:

7 failed, 224 passed, 2 skipped, 73 deselected

All seven are red-capable; none is a presence check. Both directions are pinned, so a
hardcoded constant fails too, and test_string_comparison_is_not_a_substitute_for_profile_exact
also kills the mutant that implements profile_exact as the naive string comparison.

Restored: just check && just test — 231 passed, 2 skipped. mkdocs build --strict
clean.

Not in this change

Helper.get_events drops unknown flight-event types and returns a dict that looks
complete — the same defect a layer up, filed as #63. Helper._warn_absent_once and
_warn_on_profile_drift are weaker variants of it. One logical change per commit, and
#63 needs its own decision about the return shape.

[CameronBrooks11]

Updated after independent review

The reviewer verified every claim in the original body (gate counts, the red-capable demonstration, the suppressed-logger reproduction, the RC trap, and that test_api_surface.py freezes module __all__ only) and then found four things worth fixing. All are addressed in 18f48bc.

1. The fix was undiscoverable. Four docs still taught the log warning as the mechanism — docs/guides/multi-version.md being the page a caller reads for exactly this question. Shipping the surface while every doc describes the log line leaves the defect in place for anyone not reading source. multi-version.md now shows the flag with a worked example and states why the string comparison is not equivalent; docs/index.md, README.md and docs/maintainers.md rows updated.

2. The docstring did not render. mkdocs.yml sets docstring_style: sphinx; my RST definition-list block was passed through as description text and Markdown collapsed it into a single run-on <p>. --strict does not catch this. Converted to :ivar: — the style helper.py and jiterator.py already use — and verified against the built HTML that it now renders as a table and the collapsed-paragraph form is gone.

3. SimulationPool had the identical defect. It builds an OpenRocketInstance to validate the jar and fire the warning once in the parent, then discarded it — so a pool caller, on the monte-carlo path where stale enums matter most, still had only the log line. It keeps the verdict now. Two tests added, both shown to fail without the change.

4. A comment claimed a demonstration the test did not make. The RC case's comment said it showed why or_version != profile.version_string is wrong, but the test never performed that comparison — prose asserting something with no executable assertion behind it. It is now its own test that actually asserts the trap.

Also: PR title moved from fix: to feat: (it adds public API, and the CHANGELOG files it under Added), and the CHANGELOG reference style now matches the file's bare (#n) convention instead of introducing a stranded link definition.

Gate after the changes: just check && just test — 231 passed, 2 skipped (was 228; three new tests). mkdocs build --strict clean.

Follow-ups filed separately rather than folded in here — the reviewer found other log-only sites, notably Helper.get_events dropping unknown event types from the returned dict with only a log line.