[vscode-lm] Add missing model families and correct stale context limits - #19

What is the problem?

The built-in table of VS Code Language Model models had drifted from what the API actually offers. Several model families available through the VS Code LM provider were missing entirely, and some context-window figures for models that were listed had gone stale, so the extension reasoned about budgets that no longer matched reality.

How does this PR solve the problem?

It adds the missing model families and refreshes the stale context budgets in the shared model metadata, then updates the type, provider, and hook tests that assert against that table. The new figures come from a probe run against the live VS Code LM API rather than from vendor marketing pages, so the table reflects what the extension will actually be granted at runtime.

Where the probe could only establish a lower bound rather than an exact ceiling, the conservative lower bound was used instead of guessing upward. That is a deliberate trade-off: under-reporting a budget is safe, while over-reporting causes request failures. A small number of figures remain unverified and are documented as such rather than being silently presented as measured.

The extension version is also bumped from 3.53.5 to 3.53.6 so this model-table change ships under its own version. Only the extension manifest carries the version; every other consumer derives it, so no lockfile, package, or changelog updates were needed.

Companion documentation and probe evidence: https://github.com/simurg79/roo-vault/pull/648

How did you test the PR?

• Scoped suites covering the changed areas: 98 tests passing.
• packages/types full suite: 85 passing.
• Provider suite: 1255 passing, 1 skipped.
• Version bump: 4 manifest-integrity checks plus the packages/build suite, all passing with no failures.
• Build: the packages/types package build plus type-checks across 3 packages, and ESLint: passing.
• Commit hook lint: 10/10 packages passing; type-checks 10/10 passing on the version-bump commit.
• Prettier check surfaced only CRLF line-ending noise with zero content difference; the one genuine cosmetic issue found was fixed.

The broader settings suite result is: 32 passing, 235 failing, 6 skipped. Only one representative Vertex failure group (4 passing, 1 failing) was re-run on the base commit and reproduced there; that specific failure is therefore pre-existing. The remaining failures were not established as pre-existing — no evidence supports generalising from that single sample. This PR does not claim a green full suite.

Deferred Items

Non-blocking review findings tracked separately; none change behaviour in this PR.

• [ENHANCEMENT] Tidy a redundant check and an over-stated comment in the model table tests #20 — remove a redundant check in the model table tests and soften an over-stated pre-existing comment.
• https://github.com/simurg79/roo-vault/issues/649 — the measuring tool counts unrecognised failures as successes; the affected figure here was corrected by hand, so the published values stand.
• https://github.com/simurg79/roo-vault/issues/650 — confirm the four limits the probe left unverified (one lower bound, vendor cross-check, long-conversation behaviour, one image-support unknown).