feat: add repeat wildcard flag for greedy input matching

[rowan-stein]

User Request

The E2E pipeline test in agynio/agents-orchestrator (PR #56) fails because Codex SDK sends a variable number of context messages (environment context, developer instructions, AGENTS.md) before the actual user prompt. The simple-hello test suite expects a fixed number of input items, causing sequence_exhausted errors.

Specification

Problem

The matchInput() function in src/lib/responses/matching.ts does strict 1:1 positional matching between test sequence items and actual input items. When Codex sends [env_context, "hi"] (2 items) but the test defines only ["hi"] (1 input item), the boundary check expectedItems.length === input.length fails.

Solution: Add repeat flag to wildcard message items

Add a repeat: boolean optional field to StoredMessageContentSchema. When repeat: true is set on a wildcard item (with any_role: true AND any_content: true), the matching algorithm treats it as matching 1 or more consecutive input items of any type, greedily consuming items until the next expected item matches.

Changes Required

1. src/lib/responses/matching.ts

• Add repeat: z.boolean().optional() to StoredMessageContentSchema
• Update matchInput() to handle repeat wildcards using greedy backtracking:
  • Before finding the output boundary, collect input segments with their repeat flag
  • For non-repeat segments: match exactly 1 input item (current behavior)
  • For repeat segments: try consuming N items (greedy, max→1), then recursively match remaining segments against remaining input
  • Repeat wildcard must consume at least 1 item

2. src/lib/responses/__tests__/matching.test.ts

Add tests for:

• Repeat wildcard absorbing multiple input items before exact match
• Repeat wildcard absorbing single input item
• Repeat wildcard requiring at least 1 match (fails with 0)
• Mixed repeat + non-repeat segments

3. Create simple-hello test suite

After the matching logic change, create the test via API:

{
  "name": "simple-hello",
  "items": [
    {"type": "message", "content": {"role": "user", "content": "ignored", "any_role": true, "any_content": true, "repeat": true}},
    {"type": "message", "content": {"role": "user", "content": "hi"}},
    {"type": "message", "content": {"role": "assistant", "content": "Hi! How are you?"}}
  ]
}

This absorbs Codex's variable-length context prefix, matches the user prompt "hi", and returns "Hi! How are you?".

Context

• Related: agynio/agents-orchestrator#54 (E2E pipeline test)
• The simple-hello model name is used via MODEL_OVERRIDE=simple-hello in agynd