Pin the real GC ref-count contract + correct the GC reference docs

[CoreyRDean]

Non-technical summary

The language reference's garbage-collection section (shipped in PR #67) made two claims that don't match how the runtime actually behaves, and cited a test that didn't prove them. This PR corrects the docs to the empirically-verified behavior and expands the cited test so it genuinely demonstrates the contract. Fixes a shipped doc-correctness defect (my own from #67) and turns the GC test from two trivial asserts into a real contract.

Technical summary

Verified against the live runtime (scratch .bb probes), then pinned:

• RefCount tracks the object's lifetime binding, not every alias — Local b.T = a does not raise RefCount(a) (stays 1); field assignments aren't counted either.
• A reference cycle is collected, not leaked — a\other=b : b\other=a (function locals) → after scope exit First T = Null, Each T count 0. (The field back-references add no counts, so the cycle holds nothing alive.)
• A field-only-referenced object stays alive while its owner is reachable (count 2; container\other\tag readable).

Changes:

• tests/GarbageCollectionTest.bb: keep testRef/testRel; add testAliasDoesNotIncrementRefCount, testReferenceCycleIsCollected, testFieldReferencedObjectStaysAlive (one Type per concern to isolate per-type object lists).
• help/language/lang_ref_modern.html: correct the two inaccurate claims and repoint the citation at the now-demonstrating test.
• No src/ change — this pins current behavior. The deeper reference_map-vs-obj->ref_cnt mechanism question (should aliases/fields be counted?) is flagged as a separate, deferred investigation; this PR does not change runtime semantics.

Acceptance criteria + results

• New Test blocks pass under blitzcc -t, asserting empirically-captured values (single=1, alias stays 1, cycle collected → count 0 / First = Null, field-ref object survives → count 2).
• Doc no longer claims RefCount counts all aliases, no longer claims cycles leak; the GarbageCollectionTest.bb citation is now accurate.
• test.bat green; corpus sweep unaffected; no src/ change.

Trade-offs, deferred follow-ups

• Rejected an idealized ref-count suite (aliasing→RefCount=2): the runtime doesn't behave that way; pinned reality instead (the empirical-capture pattern).
• Deferred (real open question): whether the reference_map ref-count should track aliases/field references (the reference_map vs older obj->ref_cnt distinction) — a potential GC-soundness/design question needing a dedicated investigation, not a rushed runtime change.
• Other deferred recon pitches this iteration: restore SoundPan/ChannelPan on the OpenAL backend (corpus Class B + shipped-doc gap); a TCP-socket UAF in bbsockets.cpp; contextual-keyword Phase 2.

🤖 Generated with Claude Code