laravel/ai: migrate taint coverage from class-redeclaration stubs to handler-based table

[alies-dev]

Follow-up to #937 / #484.

Problem

stubs/integrations/laravel-ai/ currently redeclares ~10 classes/traits/interfaces (AnonymousAgent, TextResponse, AgentResponse, StreamableAgentResponse, Tools/Request, Tools/SimilaritySearch, Promptable, Files/Document, Files/Image, …) purely to attach @psalm-taint-sink/@psalm-taint-source annotations.

Every redeclaration triggers CLAUDE.local.md's documented footgun: Psalm's ClassLikeNodeScanner resets class_implements / parent_interfaces and re-populates from the stub's own clauses. The Round-2 fix in #937 mitigates this by copying full implements/extends/use clauses plus blocking direct methods, but the cost is:

• Stub bodies drift on every laravel/ai release (signatures, default values, new properties, renamed methods).
• The Round-2 hybrid leaves provider-internal methods (getTimeout, withModelFailover, assertPrompted, fake, …) implicitly inherited from the reflected source — works today, but quietly breaks if the next release inlines them on the class.
• Drift detection currently depends on the application-test pass in tests/Application/laravel-test.sh (composer require laravel/ai step added in Add taint analysis stubs for laravel/ai (LLM prompt injection) #937). That catches signature breaks at analyzer-time but only on the CI matrix that runs the integration job.

Proposed approach ("path c")

Drop the redeclarations entirely. Move every annotation into a handler-driven table:

1. Psalm\\LaravelPlugin\\Handlers\\Ai\\LlmPromptSinkHandler subscribing to AfterMethodCallAnalysisEvent / AfterFunctionCallAnalysisEvent.
2. A static FQN+param-name table — same shape as Handlers\\Validation\\ValidationTaintHandler's sink list — maps Laravel\\Ai\\Contracts\\Agent::prompt#1 → llm_prompt, Laravel\\Ai\\Tools\\SimilaritySearch::withDescription#1 → llm_prompt, etc.
3. Property reads keep the existing LlmOutputTaintHandler (already handler-driven for \$response->text).
4. Return-value sinks (Tool::description(), Agent::instructions() — the inert @psalm-taint-sink llm_prompt return that we dropped in Add taint analysis stubs for laravel/ai (LLM prompt injection) #937) become expressible: the handler sees the called method, looks up its FQN in the table, and routes the return type into a synthetic taint sink.

Benefits

• Zero class redeclarations → no stripped metadata, no verbatim copy obligation per CLAUDE.local.md.
• Drift cost evaporates: the only thing that breaks on a laravel/ai release is a missing/renamed FQN entry in the table — fast to detect with a unit test asserting every entry resolves to a real method.
• Return-sink coverage lands at the same time, closing the Tool::description() / Agent::instructions() gap.
• Same registration gate as today (isInstalledAndSatisfies('laravel/ai', '^0.6')).

Estimated scope

~3h. Mostly: handler + table + unit test that walks the table against vendor/laravel/ai/ via reflection. Stubs deleted except helpers.phpstub (taint sink on a free function — already minimal).

Out of scope (separate follow-ups)

• Tool poisoning detector (non-constant Tool::description() return) — needs a different handler shape.
• Lethal Trifecta architectural issue.
• Extending the same table-based approach to prism-php-prism, openai-php-client, anthropic-ai-sdk, neuron-core-neuron-ai, llphant-llphant.

cc #484