fix: publish the real context window; refresh rates; purge the stale 8,192

[Brian Krabach (bkrabach)]

What changed

get_info() published no context_window and no max_output_tokens, so context-simple's _calculate_budget fell through to its conservative fallback on every Gemini session: a 200,000-token budget (300,000 where a bundle set one) against models that accept 1,048,576. Roughly 70% of the window went unused, silently — an undersized budget does not error, it just compacts early.

The window was never unknown. list_models() has always read input_token_limit off the live API; the number simply never reached the budget path.

• New _capabilities.py: per-model context_window / max_output_tokens, with longest-prefix matching so dated aliases resolve to their base model, and a documented default for ids newer than this file. list_models() remains the live source of truth; this answers only the synchronous question get_info() cannot make a network call to answer.
• get_info() now publishes both keys context-simple reads, for self.default_model — it previously reported a hardcoded "gemini-3.7-flash" no matter what the instance was configured with.
• The 8,192 output cap is gone from source and README. It predates every model this provider serves; all current models accept 65,536. list_models() also stopped clamping the API's own reported limit to it.

Costs re-verified 2026-09-15 against the pricing page (last updated 2026-09-11)

• gemini-3.7-flash, the provider's own default model, was missing from the rate table entirely — every call on it reported cost None.
• gemini-3.8-flash (GA 2026-09-02) and gemini-3.5-flash-lite added.
• gemini-3.6-flash carried 1.50/7.50 — the post-promo rate — as if current, over-reporting by 2x.
• 3.6/3.7/3.8 Flash are on promotional pricing through 2026-12-31, after which both rates double. Those entries now carry promo_until plus a post_promo block and compute_cost() selects by date, injectable via today= so both sides of the boundary are testable. A hardcoded promo rate would have silently under-reported by 2x from 2027-01-01 — invisible, because a wrong cost looks exactly like a right one.
• gemini-3-pro-preview removed: Google shut it down 2026-03-09. It has no published rate, so any number here would be invented; None is the honest answer.

How to verify

New test_published_context_window.py (21 tests) pins the contract at the seam that consumes it. Existing cost/config tests updated to the new rates and limits.

330 passed

Two pre-existing failures/lints are untouched and unrelated, both present on main: test_image_vision_integration_with_real_api needs a live GOOGLE_API_KEY, and ruff F401 in test_streaming.py.

Breaking changes

Cost lookups for gemini-3-pro-preview now return None rather than a rate (the model is shut down). Reported costs for gemini-3.6-flash drop 2x to the actual current promotional rate.

Coordinated five-repo change

max_tokens in context-simple was documented as "Maximum context size" but implemented as a fallback consulted only when a provider published no context window. Orchestrators always pass a provider, so the knob was silently dead in production. It is now a real cap (default None = no cap); a new max_tokens_fallback (default 200,000) took over the fallback role.

Merge order matters

Merge FIRST (independent of each other, any order):

1. amplifier-module-context-simple — feat/max-tokens-cap
2. amplifier-module-provider-gemini — feat/publish-context-window
3. amplifier-app-cli — feat/session-module-config-overrides

Merge AFTER those three:
4. amplifier-foundation — chore/drop-dead-max-tokens
5. amplifier-bundle-attractor — chore/drop-dead-max-tokens

Rationale: the two config repos delete max_tokens lines that only become safe-to-delete once context-simple's new semantics are in.

Cross-repo verification

DTU instance context-overflow-fix-20260915, all 8 target behaviors PASS:

• A Gemini session's effective budget goes from 200,000 (the old fallback) to 1,011,712 (its real published window) — a 5.1x increase.
• max_tokens: 500000 caps to exactly 500,000.
• A cap above the model window is a no-op.
• settings.yaml overrides.context-simple.config now reaches session.context through the CLI's own resolve_bundle_config.

Repo	Result
context-simple	166 passed, 1 xfailed
provider-gemini	330 passed; 1 pre-existing failure (test_image_vision_integration_with_real_api, needs a live GOOGLE_API_KEY, fails on main too)
app-cli	2335 passed, 2 skipped, 1 xfailed
amplifier-foundation	1938 passed, 3 skipped; 1 pre-existing failure (test_grpc_adapter_main.py::TestVerifyModuleType::test_non_isinstance_object_with_mount_passes, verified failing on main before the change)
amplifier-bundle-attractor	268 passed, 2 skipped

[Brian Krabach (bkrabach)]

The five PRs in this coordinated change

Merge FIRST — independent of each other, any order:

1. context-simple — feat: max_tokens becomes a real cap; max_tokens_fallback takes over the fallback amplifier-module-context-simple#42
2. provider-gemini — fix: publish the real context window; refresh rates; purge the stale 8,192 #48
3. app-cli — fix: config overrides reach session.context and session.orchestrator amplifier-app-cli#342

Merge AFTER those three:
4. amplifier-foundation — microsoft/amplifier-foundation#388
5. amplifier-bundle-attractor — microsoft/amplifier-bundle-attractor#359

The two config repos (4, 5) delete max_tokens lines that only become safe-to-delete once context-simple's new cap semantics (1) are in.