fix(parsers): MermaidParser no longer silently drops edge lines, plus metadata pass

[thorwhalen]

The bug

MermaidParser discarded any edge line it could not match, without a word. A
live gpt-4o-mini call produced this:

flowchart TD
    start([Start]) --> check_user{Is user authenticated?}
    check_user --|Yes|--> show_dashboard([Show Dashboard])
    check_user --|No|--> show_login([Show Login])
    show_dashboard --> end([End])
    show_login --> end([End])

check_user --|Yes|--> show_dashboard is invalid Mermaid — the valid form is
check_user -->|Yes| show_dashboard. The parser dropped both of those lines,
and the first line fared no better: an edge line carrying inline node
definitions failed the edge patterns, fell through to the node parser, which
matched whichever shape came first in its pattern list and threw the rest away.

Net result — the parser returned this, reporting nothing wrong:

	nodes	DECISION nodes	edges
what the diagram says	5	1 (check_user)	5
what the parser returned	4	0	2

check_user{Is user authenticated?} was gone. show_dashboard and
show_login were typed START. A diagram that should contain a decision
contained none, and nothing anywhere said so. This is how model drift becomes a
silently wrong diagram rather than an error.

This is not hypothetical drift: it is exactly why
test_llm_converter_real_api_with_decision fails today
(assert len(decision_nodes) >= 1 → assert 0 >= 1) on any machine with
OPENAI_API_KEY exported. The A[Start] --> B[End] idiom — the most common
Mermaid form there is — was equally affected; the pre-existing
test_llm_converter_code_block_cleanup used it and passed anyway, because it
only asserted validate(), which an under-populated diagram satisfies.

The fix

ij/parsers/mermaid.py:

• Edge endpoints are resolved individually, so inline node shapes on an
  edge line are kept and typed correctly.
• Non-canonical link spellings are normalized to the canonical
  ARROW|label| form: the invalid -- |label| --> drift, plus the valid
  Mermaid alternatives a -- label --> b and a -. label .-> b which were
  also being dropped.
• Chained edge lines (a --> b --> c) yield one edge per link. Previously
  only the first link survived.
• Nothing is dropped mutely. An uninterpretable line is recorded in
  MermaidParser.unparsed_lines and raised as a MermaidParseWarning; the new
  keyword-only strict=True turns it into the ValueError that parse's
  docstring has always promised but never raised.
• Valid Mermaid must not trip that warning, so comments (%%),
  subgraph/end/direction/style/classDef/click/linkStyle
  directives, bare node ids, and the legacy graph keyword are now recognized
  rather than falling through.
• Node shapes are a single table driving both the node and endpoint patterns,
  instead of two hand-maintained pattern lists.

ij/converters/llm_converter.py: convert, refine and
convert_with_examples all route through a new _parse_mermaid, which refuses
output holding no valid diagram (with the raw model output in the message)
instead of returning an empty one. MermaidParseWarning is exported from ij
so a caller can escalate it with warnings.simplefilter("error", ...).

Confirmed against the live API: with this fix,
test_llm_converter_real_api_with_decision passes against real gpt-4o-mini —
the root cause was the parser, not the model.

Tests

	before	after
gate (priv test-dependents ij)	123 passed, 1 failed	131 passed, 2 skipped
CI dry run (wads doctest flags)	123 passed, 1 failed	131 passed, 2 skipped

Eight new hermetic tests (no network). Six in tests/test_parsers.py pin:
inline node definitions surviving; the --|label|--> normalization (asserted
under simplefilter("error"), so a dropped line fails the test); the
-- label --> / -. label .-> forms; chained edges; the warning and
unparsed_lines on genuinely unreadable input; strict=True raising; and a
guard that a valid diagram full of comments, subgraphs and style directives
produces no warning at all. Two in tests/test_llm_converter.py pin the
end-to-end drift case (the captured model output above, asserting the DECISION
node survives — the hermetic stand-in for the live test) and the refusal of
unusable output.

Live-API tests are now opt-in

Both real-API tests were gated on mere presence of OPENAI_API_KEY, so any
developer or agent with a key exported silently made non-hermetic,
network-dependent, cost-incurring calls. They now require an explicit
IJ_RUN_REAL_API_TESTS. CI behaviour is unchanged — it sets neither variable,
so both were and remain skipped there.

Metadata pass

• license: the [project.license] table with the non-SPDX value "mit"
  becomes the inline SPDX string license = "MIT". Verified in the built
  wheel: License-Expression: MIT, Metadata-Version: 2.4. No
  License :: OSI Approved :: trove classifier is added alongside it — PEP 639
  rejects that combination.
• authors: was [], so PyPI showed no author.
• classifiers: none were declared. The 3.10/3.12 pair mirrors the CI matrix.
• .editorconfig: added, byte-identical to the fleet template.
• ci.yml: dropped two stale migration comments (no setup.cfg exists;
  PYPI_PASSWORD is provisioned and Publish has succeeded). No behavioural
  change.
• Repo metadata set out of band: homepage → https://i2mint.github.io/ij/, plus
  the seven topics mirroring the pyproject keywords. Description was already
  correct and was left alone.

Deliberately left out of scope

• A validate-then-retry loop in LLMConverter. The parser fix plus a hard
  failure on unusable output covers the observed drift; a retry loop is a
  bigger design question (retry budget, prompt repair, cost) and belongs in its
  own PR.
• Mermaid open links (a --- b, a -.- b) are still unsupported — but they
  now warn instead of vanishing.
• ij.__version__ is "0.2.0" while pyproject and PyPI are at 0.1.5.
  Pre-existing drift, untouched.
• Package doctests never run. testpaths = ["tests"] means the wads CI
  invocation collects only tests/, so ij/'s doctests are not executed
  anywhere. Two are broken or non-hermetic:
  LLMConverter.refine's raises NameError: name 'converter' is not defined,
  and ij/export/image.py's writes a diagram.png into the repo root. Changing
  testpaths would change what the gate collects, so it is left for a separate
  decision.

https://claude.ai/code/session_01VipiLaG4xy7WctqY9w2475