fix(schema): ModelAlias throws AliasMaxDepthError for a chain that settles exactly at the limit — 9-hop chain rejected, effective limit is 8 #399

Problem

ModelAlias.resolveRef throws AliasMaxDepthError for an alias chain that successfully settles exactly on the depth limit. The post-loop check cannot distinguish "the loop settled and broke out" from "the loop ran out of depth", so a legitimate chain one hop shorter than the apparent limit already fails.

Evidence (verified 2026-10-24 on main 58dd3fb)

File: packages/schema/src/model-alias.ts

• MAX_DEPTH = 10 (line 32); loop is while (depth < MAX_DEPTH) (line 61), incrementing depth at the top of every iteration — including the final iteration that merely confirms the settled provider/model and breaks.
• After the loop, line 153 unconditionally throws when depth >= MAX_DEPTH, even when the loop exited via break with a fully resolved reference.

Reproduction (Bun, importing packages/core/src/model-alias.ts; global model aliases a0 -> a1 -> ... -> a(n-1) -> openai/gpt-4o, i.e. n expansion hops):

A 9-hop chain needs 9 expansion iterations plus 1 settling iteration = depth 10, hits the post-loop check, and is rejected despite having resolved. The effective usable limit is therefore 8 hops, not the documented-by-constant 10, and the error message misreports a resolved chain as a runaway one.

Why it matters

Alias chains are how users layer naming (team alias -> provider alias -> concrete model). A chain that is within the intended limit fails with an error that claims the opposite, and the failure boundary (8 usable hops) matches neither the constant nor any documented limit, so users cannot predict whether a config will work without counting loop iterations including the invisible settling pass.

Acceptance criteria

• A chain that settles within MAX_DEPTH total loop iterations resolves instead of throwing — in particular the verified 9-hop chain above returns openai/gpt-4o.
• A chain that genuinely never settles within the limit (cycle-free but longer than the limit, and cycles longer than the visited-set catches) still throws AliasMaxDepthError; cycle detection (AliasCycleError) is unchanged.
• The fix distinguishes settled-vs-exhausted loop exit (e.g. track settlement explicitly) rather than just raising MAX_DEPTH, so the boundary stays exactly at the constant.
• New tests in packages/core/test/model-alias.test.ts cover the boundary: the largest chain that should resolve, and the smallest that should throw. Existing tests (including test 13, chained expansion) still pass.
• Dedup checked 2026-10-24: no open or closed issue mentions AliasMaxDepth / max-depth alias behaviour.

Single-repo: ModelAlias lives only in this repo's schema/core packages.