feat(provider): add native Google Gemini API adapter (Phase 4)

Adds internal/provider/google, the second native cloud adapter (following
internal/provider/anthropic's pattern) on top of Phase 1's provider-neutral
tool-use loop, wired to a Docker-sandboxed NativeRunner under agent.type:
"google" -- a separate execution path and budget bucket from the existing
CLI-subprocess "gemini" ContainerRunner, which is untouched.

Wire-format research (the highest-risk part of this adapter): Gemini's
multi-turn function-calling shape was resolved by cross-referencing the
REST API reference's own generateContent example against the go-genai SDK's
struct tags on GitHub -- both agree on functionCall/functionResponse parts
keyed by "name" (with an optional "id" for round-tripping ToolCall.ID),
with the response fed back inside a "user"-role Content (Gemini has no
tool/function role, mirroring Anthropic's lack of one). A separate fetched
source (the function-calling guide page) was deliberately discarded as a
reference for this shape -- it documents a different, newer "Interactions
API" whose call_id/type:"function_result" structure doesn't fit the
contents/parts/candidates shape used everywhere else.

- internal/provider/google: request/response translation, systemInstruction
  handling, role mapping (assistant->model, tool-results->user role),
  per-model-prefix pricing table (2.5 Pro/Flash/Flash-Lite, 2.0, 1.5 tiers)
- internal/retry: IsRateLimitError additively extended for RESOURCE_EXHAUSTED
- internal/config: RunnersConfig.Google/GoogleEnabled()
- internal/cli/serve.go, run.go: runners["google"] construction mirroring
  the Anthropic wiring exactly (Docker sandbox default)
- docs/api-keys-setup.md: Google marked wired-up, budget-bucket/disable/
  verify guidance added matching the Anthropic section

go build/vet/test -race all pass. No live Gemini API key available in this
environment; verified via fake-httptest-server adapter tests (plain text,
tool-use round-trip, multi-turn tool-result, rate-limit error matching) plus
a Pool/NativeRunner routing test. Live E2E is a follow-up once a key is
configured, same as Phase 2's Anthropic adapter.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V1moSNCJRcP6kykA4tyUSs

1 files changed, 60 insertions, 0 deletions

diff --git a/internal/provider/google/pricing.go b/internal/provider/google/pricing.go
new file mode 100644
index 0000000..55ea66f
--- /dev/null
+++ b/internal/provider/google/pricing.go
@@ -0,0 +1,60 @@
+package google
+
+import "strings"
+
+// modelPrice is USD cost per million tokens for one model (or model family
+// prefix).
+type modelPrice struct {
+	InputPerMTok  float64
+	OutputPerMTok float64
+}
+
+// pricingTable maps a model-name *prefix* to its per-million-token pricing.
+// Keyed by prefix (rather than exact model string) so dated/versioned model
+// IDs (e.g. "gemini-2.5-flash-preview-09-2025") still resolve without
+// needing an entry per snapshot. Edit this table as Google's pricing or
+// model lineup changes — it intentionally has no other logic attached.
+//
+// Prices below are cached from training/research as of mid-2026 and are not
+// fetched live; verify against https://ai.google.dev/gemini-api/docs/pricing
+// before relying on them for real billing decisions. Two caveats specific to
+// Gemini pricing that this flat per-model table does NOT model:
+//   - Some tiers (notably 1.5 Pro and some 2.5 Pro contexts) charge a higher
+//     rate once the prompt exceeds 128K/200K tokens; the price below is the
+//     lower/base-context rate.
+//   - Some Flash-tier models charge a different (higher) output rate for
+//     "thinking"/reasoning tokens vs. plain output tokens; the price below
+//     is the standard output rate, not the thinking-token rate.
+var pricingTable = map[string]modelPrice{
+	// Pro tier (most capable, most expensive).
+	"gemini-2.5-pro": {InputPerMTok: 1.25, OutputPerMTok: 10.00},
+	"gemini-1.5-pro": {InputPerMTok: 1.25, OutputPerMTok: 5.00},
+
+	// Flash tier (balanced cost/capability; default for most tasks).
+	"gemini-2.5-flash-lite": {InputPerMTok: 0.10, OutputPerMTok: 0.40},
+	"gemini-2.5-flash":      {InputPerMTok: 0.30, OutputPerMTok: 2.50},
+	"gemini-2.0-flash-lite": {InputPerMTok: 0.075, OutputPerMTok: 0.30},
+	"gemini-2.0-flash":      {InputPerMTok: 0.10, OutputPerMTok: 0.40},
+	"gemini-1.5-flash-8b":   {InputPerMTok: 0.0375, OutputPerMTok: 0.15},
+	"gemini-1.5-flash":      {InputPerMTok: 0.075, OutputPerMTok: 0.30},
+}
+
+// defaultPricing is used for unrecognized models rather than erroring —
+// Flash-tier pricing is a reasonable default since it's the most commonly
+// used tier for the harness's escalation ladder.
+var defaultPricing = modelPrice{InputPerMTok: 0.30, OutputPerMTok: 2.50}
+
+// pricingFor returns the per-million-token input/output pricing for model,
+// matching the longest pricingTable prefix that model starts with. Falls
+// back to defaultPricing for unrecognized models.
+func pricingFor(model string) (inputPerMTok, outputPerMTok float64) {
+	price := defaultPricing
+	bestLen := -1
+	for prefix, p := range pricingTable {
+		if len(prefix) > bestLen && strings.HasPrefix(model, prefix) {
+			bestLen = len(prefix)
+			price = p
+		}
+	}
+	return price.InputPerMTok, price.OutputPerMTok
+}