feat(proxy): use tenant catalog token limits for VS Code models

[bohdanmaliar]

Summary

codemie proxy connect vscode writes the languageModels config for VS Code from a hardcoded capability table. Those token limits were static, so models whose limits differ per tenant got the wrong maxInputTokens / maxOutputTokens. This PR reads the limits from the tenant model catalog and uses the table only as a fallback.

Changes

• Catalog parsing (tenant-catalog.ts): TenantModelDescriptor now carries maxInputTokens and maxOutputTokens from max_input_tokens / max_output_tokens. Values are kept only when they are finite numbers > 0. Zero, negative, string, null and missing values are dropped.
• Limit resolution (vscode-models.ts): new resolveVsCodeTokenLimits(entry, descriptor).
  • Output limit: catalog value, else the table entry (or the default for untabled models).
  • Input limit: catalog input minus the resolved output limit. The catalog value is the full context window, whereas VS Code's maxInputTokens is a prompt budget, so input plus output has to fit.
  • If the catalog input is missing, or the subtraction is not positive, the entry's own maxInputTokens is used.
• Config writer (vscode.ts): buildManagedModel takes the descriptor and applies the resolved limits to both table-matched and default (untabled) models.
• sso.http-client.ts: small type tweak (LlmModel) so the catalog type can be reused.
• docs/COMMANDS.md: documents the resolution order (tenant catalog, then built-in table, then defaults) and the input-minus-output rule.

Examples

Model	Catalog	Result (in / out)
claude-4-5-sonnet (tabled)	input 200000	136000 / 64000
gpt-6-sol (untabled)	input 922000	913808 / 8192
any	input 200000, output 16000	184000 / 16000
claude-4-5-sonnet	input 64000	entry input / 64000 (fallback, since 64000 - 64000 is not positive)
claude-4-5-sonnet	no limits	table values unchanged

Tests

Added unit tests for catalog parsing (valid and invalid values), resolveVsCodeTokenLimits (all cases above) and an end-to-end config-write test covering both tabled and untabled models.

The docs/superpowers/tasks/2026-10-02-vscode-model-token-limits/ files are planning artifacts from the SDLC run that produced this change.