zai, zhipuai: add missing limit.input for GLM-5.x models

[winstern1998-commits]

Summary

Adds the missing limit.input for the GLM-5.x family, superseding #3739 (closed as stale) and incorporating the review bot's action items from that PR.

Zhipu does not publish a separate max-input figure; the official docs define the context window as the combined total of input and output tokens (concept-param), so limit.input is derived as context - output — the same convention already used for OpenAI entries in models/ (e.g. gpt-5.5: 1_050_000 = 922_000 + 128_000).

Per-model values (official context/output from docs.bigmodel.cn model pages, cross-checked with docs.z.ai where available):

Model	context	output	input
GLM-5	204_800	131_072	73_728
GLM-5.1 / GLM-5-turbo / GLM-5v-turbo	200_000	131_072	68_928
GLM-5.2 / GLM-5.3 / GLM-5.3-Flash	1_000_000	131_072	868_928

Changes

• models/zhipuai/glm-5* (7 files): canonical metadata gets limit.input. This is the primary fix — zai and coding-plan entries resolve these via base_model, which is why editing providers/* alone (as zai, zhipuai: add missing limit.input for GLM models #3739 did) never propagated.
• providers/{zai,zhipuai,zai-coding-plan,zhipuai-coding-plan} (9 files): first-party entries that carry their own [limit] block.
• Third-party base_model entries (24 files): entries whose own context is smaller than the inherited input get an explicit input (= own context − own output) so they don't end up with input > context after this change. Degenerate entries (output >= context, where the subtraction would go negative) use base_model_omit = ["limit.input"] instead — the same pattern the sync tooling itself emits (baseModelOmit in sync/providers/openrouter.ts).
• sync.test.ts: the LLM Gateway factoring test now expects base_model_omit: ["limit.input"] since the base model gained limit.input — expected baseModelOmit() behavior, no product change.

ambient/zai-org/GLM-5.1-FP8 intentionally keeps the inherited input (68_928 ≤ context 202_752).

Downstream motivation

opencode's compaction.reserved is silently ignored for models without limit.input because usable() only applies the reserved buffer on the limit.input branch (anomalyco/opencode#38835). With limit.input present in the catalog, compaction.reserved takes effect with no opencode-side change.

Validation

• bun run validate exits 0
• bun test: all affected tests pass (8 pre-existing failures unrelated to this change, confirmed via clean-baseline stash on WSL)
• Full-catalog invariant scan (via generateCatalog): 570 GLM entries across 203 providers, zero input > context violations; all first-party ^glm-5 entries have input; e.g. zai/glm-5.2 resolves to {context: 1000000, input: 868928, output: 131072}

A follow-up PR could extend limit.input to the GLM-4.x family (10 files, same shape).

[github-actions]

No actionable findings.