Superseded: Add SayGM provider

[markdavison]

Superseded by #4625, which incorporates the automated review feedback in a single clean commit.

[github-actions]

Action items

• [high] [violation] providers/saygm/models/gpt-5.4.toml:2 - Check: base_model_omit applies after merge and must not drop provider-authored fields that should remain. Why: Every SayGM model sets base_model_omit = ["cost"] while also authoring [cost]. Generator merge is base → overrides, then omit, so omit deletes the provider’s own cost from the resolved catalog (lab metadata under models/ has no cost to strip). Published SayGM prices never appear in generated output. Action: Remove base_model_omit = ["cost"] from all providers/saygm/models/*.toml entries (and stop writing it in the synchroniser/tests).

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/saygm.ts:1658 - Check: Update-only sync must retain reviewed models through incomplete/outage responses. Why: buildSaygmModel throws when price_range.ceiling is missing. Hourly sync will fail the whole provider job if any already-cataloged ID is still listed without a published-retail ceiling (or with a stripped price_range), instead of keeping the hand-reviewed TOML. That conflicts with the stated deleteMissing: false / retain-through-outages design and with peers such as Ofox that fall back to authored cost. Action: When ceiling pricing is absent, preserve the existing authored cost (and other fields) and skip the price update—or skip that source row—rather than throwing.
• [medium] [possible mistake] providers/saygm/models/claude-sonnet-4-6.toml:13 - Check: Provider cost should match this host’s published-retail ceiling and not silently understate standard Anthropic list pricing. Why: SayGM authors Sonnet 4.6 at output = 10 while first-party Anthropic and essentially every other catalog host use $15/MTok output (with $3 input). If the live ceiling is not actually $10, consumers will under-estimate cost. Action: Verify price_range.ceiling for claude-sonnet-4-6 on GET https://api.saygm.com/v1/models and either correct the TOML or cite the SayGM ceiling fields that justify $10 output.
• [medium] [possible mistake] providers/saygm/models/gpt-5.5-pro.toml:8 - Check: Long-context [[cost.tiers]] should be authored when this host’s published retail includes a context band. Why: OpenAI’s first-party gpt-5.5-pro and SayGM’s other GPT-5.x entries carry a 272_000 tier, but this file only has base input/output/cache_read. If SayGM’s ceiling exposes long-context dimensions for this ID, the static catalog and future syncs will omit a material price band. Action: Confirm the live ceiling for gpt-5.5-pro; add the tier when present, or note in the leading comment that SayGM does not publish long-context retail for this model.