Enable Opus 5 support, native thinking fixes, proper mapping for effort

[guidan-nick]

Thinking never rendered in OpenCode on Kiro models, and xhigh effort could not be reached. Three independent causes.

Fixes

• Reasoning was discarded. Kiro streams reasoning as reasoningContentEvent, which transformSdkStream did not handle, so every event was dropped. Now forwarded as reasoning_content deltas. Measured on opus-5: 17 events in / 0 deltas out before, 20 in / 19 out after. The one-event gap is the trailing signature/redactedContent event, skipped deliberately since it has no renderable text.
• OpenCode was never told the models can reason. It drops reasoning deltas unless the model declares reasoning: true and interleaved.field. The plugin emitted reasoning_content but declared neither. Both are now set on every -thinking entry, matching how OpenCode's own registry declares non-standard reasoning fields.
• xhigh was unreachable. budgetToEffort had no branch returning it, so with auto_effort_mapping on by default it was dead configuration — adding models to XHIGH_CAPABLE_MODELS had no effect. Confirmed on the wire after the fix: {"output_config":{"effort":"xhigh"}}, HTTP 200.

The old mechanism was prompt injection plus scraping <thinking> tags out of the answer. Opus 5 never emits those tags, so thinking was invisible there regardless of config. The scraper stays as a fallback for models that do inline reasoning.

Also

• Model list moves to plugin/model-registry.ts; thinking entries derive from effort capability so variants and budget bands cannot drift apart. xhigh only offered on opus-4.7/4.8/5 and sonnet-5.
• Adds missing claude-opus-5, claude-opus-5-thinking, claude-opus-4-7-thinking entries.
• GPT-5.6 tiers deliberately not advertised — they use reasoning.effort/reasoning.mode, not output_config.effort, so they need their own request path.
• logSdkRequest records the resolved additionalModelRequestFields; effort was previously invisible in logs and unconfirmable without patching the plugin.
• README install example no longer tells users to hand-write the model registry, which is what left thinking models without reasoning flags to begin with.

Breaking

Budget bands rescaled to Kiro's real range (1024–128000) instead of OpenCode's 32768 convention, so xhigh and max are reachable from a budget:

budget	effort
<= 16384	low
<= 32768	medium
<= 65536	high
<= 98304	xhigh
> 98304	max

A budget of 32768 now maps to medium rather than max. Anyone relying on the old bands should set an explicit effort in kiro.json.

Testing

173 tests pass, tsc --noEmit and Prettier clean. New coverage for registry flags, xhigh gating, variant/band agreement, and the reasoning stream including the live event shape. Verified end to end against a live account with provider.kiro.models removed, so the plugin's registry supplied everything.

Every behavioural claim here is wire-verified. Nothing speculative: an earlier revision of this branch added a send_thinking_config option to send Kiro's thinking block explicitly, and it was dropped because the API's acceptance of those fields was never tested and reasoning works without them.