feature: transparent model and provider switching — continue a session across a limit, price or policy boundary

Feature Description

@rubambiza's idea: when a session approaches a token or rate limit, Cortex switches the request
to a different model or provider in the background, without interrupting the user's work —
the agent keeps going instead of stalling on a 429 or a context-window error.

The same mechanism serves three other triggers:

• Cost — route to an equivalent-quality, cheaper model or provider when one is available
• Preference — a user or team pins a preferred model per task class
• Policy / compliance — a request that would violate a data-residency, provider-approval or
  region constraint is routed to a compliant provider, or refused with a clear reason rather
  than silently sent

Cortex is uniquely placed to do this: it already sits in the request path with the request body
decrypted, and already rewrites bodies (Context Guru compaction, tool-definition pruning). This
is a new plugin in that pipeline, not new architecture.

Proposed Solution

• Provider abstraction: translate a request between provider wire formats (Anthropic
  Messages ↔ OpenAI Chat Completions ↔ others) including tool definitions, tool results,
  system prompts and streaming events
• Trigger evaluation before the upstream call: projected token count vs the target model's
  window, observed rate-limit headroom, price table, policy rules
• Retry-on-limit path: a 429 or context-overflow from the primary is re-issued against the
  fallback rather than surfaced to the agent
• Preserve streaming — the client must not be able to tell a switch happened mid-turn, and
  the fix for #926 (early flush of leading thinking block) applies here too
• Credentials for each provider resolved through AuthBridge, so the agent never holds more
  than one provider's secret
• Every switch is recorded and visible. A silent model change is a correctness hazard:
  the user must be able to see, per request, which model actually answered, why it was
  chosen, and what it cost. Depends on #950
• Opt-in, off by default, with a documented list of substitutions the user has approved

Risks and open questions

• Silent quality change is the real risk. "Without interrupting your work" and "you should
  know which model answered you" are in tension. Proposal: never interrupt, always record, and
  surface a visible marker in abctl and in the session summary
• Semantics are not preserved across providers: prompt caching, thinking blocks, tool-call
  formats and stop reasons all differ. A switch mid-session can invalidate a cache the user was
  paying to build — the cost model has to account for that
• Whether policy enforcement belongs here or in the existing egress-control / guardrails plugins
• Does the agent's own context (conversation history) transfer cleanly, or does a switch
  effectively start a new upstream cache lineage?

Additional Context

Credit: @rubambiza — team brainstorming, September 2026. Post-v0.9.0 — this is the largest of the
four brainstorm items and would not fit the release window.

Foundational for #1020 (task-class routing) and #1021 (parallel dual-model comparison), which
both need the provider abstraction and the switch mechanism this issue builds.
Depends on #950 for per-model cost attribution.

Part of the post-v0.9.0 brainstorm backlog under #962.

No files or tests are named. Start by mapping the request path and existing Context Guru, egress-control or guardrails, and AuthBridge plugins, then read the constraints in #950 and the streaming concern in #926. Done requires a scoped provider abstraction and switch design that resolves the policy, cache, streaming, attribution, and visibility questions.