fix: GLM reasoning effort, [1m] request strip, and catalog windows

[xz-dev]

What

Aligns the proxy with current GLM Coding Plan contracts for reasoning effort and the official [1m] listing aliases.

1. Reasoning effort — model-specific maps (GLM-5.3 none/minimal/low → low, medium/high → high, xhigh/max → max; GLM-5.2 can disable). Shared translator covers Chat, Responses, Claude Messages, and async. Catalog is the single source of truth: Codex ?client_version and Anthropic anthropic-version advertise the official efforts.
2. [1m] request path — catalogs keep glm-5.2[1m] / glm-5.3[1m]. Shared transformRequestBody strips trailing [1m] before upstream so Chat/Responses/Claude/async stop hitting 1214 / modelCode不存在.
3. Catalog windows — bare glm-5.2 / glm-5.3 advertise 200k. Only [1m] aliases advertise 1M. Official Coding Plan 1M is opt-in via the suffix (Enable 1M Context, ZCode Connect Models).

Why

• Non-none effort previously collapsed to Anthropic thinking:{type:"enabled"}, so low and max looked identical upstream.
• Listing [1m] aliases were forwarded as modelCode, which the provider rejects (1214).
• Catalog advertised the 1M ceiling on the bare ids, so clients compacted/budgeted as if every glm-5.3 call were 1M.

Tests

Focused suites for the translator, body transformer, catalogs, and providers pass. Full bun test: 550 pass / 1 fail. The failure is pre-existing (src/integration.test.ts still loads config.test.yaml, removed in f6aa147) and is unrelated to this branch.

Notes

• No plan / async behavior change.
• README model table still lists glm-5.2/5.3 as 1M; I can follow with a docs-only commit if wanted (GPG pinentry timed out on that commit).

[TriDefender]

Hey I checked this commit and found a few issues:

Blocking issues:

• Regression: reasoning_effort: "none" now forces thinking ON for every model without an effort map (glm-4.5-air, 4.6, 4.6v, 4.7, 5, 5-turbo, 5v-turbo, 5.1, unknown ids). I re-confirmed in source: normalizeReasoningEffort returns undefined for mapless models → translateReasoning falls through to {thinking: enabled}. Master sent disabled. Live probe proved upstream honors disabled for these models — genuine regression. Fix: global fallback none/minimal + no map → {type:"disabled"}.

[TriDefender]

After careful review this PR will be closed without merging as it's considered NOT A BUG

Breaking down the PR it's obvious that 3 issues are adressed, respectivly the 1M marker, disable thinking and proxy side reasoning effort map.

Let's start with the effort map.

• Now, if you've read the original docs, it would be hard to miss that the server side handles reasoning effort already, thus it would be redundant to implement proxy side adapter.
• Then the reasoning part, as mentioned above and per test results shows, PR 30 has a serious regression. From my one-off test all models that previously behaved fine on the current branch leaked thinking, therefore this is not a "fix", it's rather the bug itself.
• Last but not the least, the 1M marker is no longer necessary in Claude Code or Codex. You may use CLAUDE_CODE_MAX_CONTEXT_TOKENS to enable 1M long context in claude code harness, for codex set it explicitly in config.toml.
• Also the [1m] marker won't be sent out anyways on CC since it's just local handling so I don't understand why you need a fix to that.