[DOCS]: fill the gaps left after #222 and rename "threshold-gated recursion"

[sromoam]

Context

Rescoped. This issue originally proposed building one consolidated threshold page. #222 already does most of that with Getting-Started/thresholds-and-metrics.md, so building it here would mean writing a competing page around an open contributor PR.

This issue is now the remainder: what #222 does not cover, plus the rename. It depends on #222 landing first.

Already covered by #222

• all four thresholds and what each gates
• threshold precedence and the rule that a comparator's own threshold is ignored in normal evaluation
• the five-category confusion matrix and the FP = FA + FD split
• recall_with_fd, both formulas, and when to use each
• a worked example end to end (numbers being corrected on that PR)

Remaining gaps

1. FN → FD raises recall; TP → FD lowers it

Two opposite moves, same vocabulary, and nothing distinguishes them anywhere in the docs:

move	numerator	denominator	default recall
TP → FD (raise match_threshold)	-1	-1	falls, or stays equal
FN → FD (what #224 changed for 1-vs-1 lists)	unchanged	-1	rises

This is worth calling out explicitly because it is genuinely easy to get backwards: a paragraph on #225 stated it inverted, and catching it needed an exhaustive check over tp0 ∈ 1..4, fn ∈ 0..3 (result: 0 cases where a TP→FD move raised recall, 6 equal, 34 fell). hungarian-matching.md now carries a corrected version of this; it belongs wherever recall_with_fd is explained.

2. The threshold=0.0 trap

A threshold of exactly 0.0 is satisfied by every score, so every pair becomes a TP and a wholly incorrect prediction reports perfect precision, recall, F1 and accuracy. #236 adds a runtime UserWarning whose message links to the threshold docs, so the docs need a section for it to land on.

It is a cliff rather than a slope, which is the part worth writing down: 0.01 classifies correctly, only 0.0 misbehaves.

3. Rename "Threshold-Gated Recursive Evaluation"

docs/docs/Advanced/threshold-gated-evaluation.md names the mechanism rather than the reader's question, and undersells itself: the page is about classification, not only recursion.

Candidate: "How Thresholds Affect Your Metrics". Alternatives: "Thresholds and Classification", or keep it as a mechanism reference and let #222's Getting-Started page be the explainer it links to.

Worth deciding deliberately, since the name becomes the link target from #236's warning message and from CHANGELOG entries.

4. Reduce duplication between the three pages

With #222 landed, hungarian-matching.md, threshold-gated-evaluation.md, and the new Getting-Started page will each partly explain thresholds. The Advanced pages should become mechanism references that link to the explainer rather than restating it.

Acceptance criteria

• The FN→FD versus TP→FD contrast appears wherever recall_with_fd is explained, with the direction for each
• The threshold=0.0 trap is documented, and feat(thresholds): warn when a threshold is exactly 0.0 #236's warning message links to it
• threshold-gated-evaluation.md is renamed, with .nav.yml and inbound links updated
• The Advanced pages link to the explainer instead of restating threshold semantics
• mkdocs build warning count does not increase (currently 9, all pre-existing)

Notes

Depends on #222. Gaps 1 and 2 came out of the #225 review; both were symptoms of threshold behavior being spread thin across three pages with no single owner.