Align list query parsing with official families

[SaladDay]

Core rejected many list and resource query strings that the official Agents API accepts or rejects differently. This batch aligns the shared list parser and query handling with owned official observations. The pinned baseline is unchanged (SDK 3.13.0 / d7c41ef / agents=v1).

Behavior

• Unknown query keys are ignored on shared-parser lists and on single-resource routes. Previously they returned 400 unsupported_parameter. Only authentication selects the tenant. Foreign and missing resources stay indistinguishable, and body validation is unchanged.
• A repeated supported list key still returns 400, now with the observed family fields:
  • Beta lists: invalid_request_error with the duplicate-field message.
  • Skills: duplicate_parameter, param limit/key.
  • Files: unchanged.
• Limits, by family:
  • Agent, Session, Item and Template lists treat 0 as 1 and values above 100 as 100.
  • Turn, Subagent and Artifact lists still reject values outside 1–100, now with code invalid_request_error.
  • Negative and non-integer Beta limits return the observed deserialization error.
  • Vault and Credential lists keep clamping, including negative values, because the pinned docstring says "clamped between 1 and 100".
  • Skills and Skill versions accept 0–100. Zero returns an empty page whose has_more reports remaining resources. Out-of-range values return integer_above_max_value / integer_below_min_value with param limit.
  • Files range errors carry a null code.
• A scalar status combined with status[] filters Vaults/Credentials by their union.
• An explicit empty Files purpose= means no filter.

Decision: the pinned Turn/Item/Template docstrings say "between 1 and 100". Clamping keeps the effective page inside that range, so Items and Templates follow the live service's clamp. Turns keep rejecting because the live service rejects.

Evidence

Campaign scan 1 sent 252 owned official requests across three families, with request IDs retained privately. It included 4 official none Sessions with 6 tiny Turns; every owned official resource was deleted and verified. The matrix, the decisions and the deferred items are recorded in contracts/agents-api/list-query-semantics.md and operation-evidence.md (entry L).

Validation

• Independent acceptance, written from the requirements only, against real Core HTTP with dedicated PostgreSQL, using raw HTTP plus the pinned SDK 3.13.0: 845 checks, 116 of them through the SDK.
  • On main, every changed row failed and every retained row passed.
  • On the candidate runtime (5e1171b), all rows passed.
  • The checks cover tenant B foreign-equals-missing, and snapshot/row fingerprints showing that no rejected request writes.
• Go handler and store tests for every matrix row. A PostgreSQL store test covers the Skills zero page. The repository's official-client acceptance passed locally.
• Server gate on this head: make -o check-web check plus Web typecheck, core-doctor, 287 client tests, 583 Web tests and the build all passed. The Playwright browser cases were not run on the server, because it has no Google Chrome; this skip was approved by the user.
• Generated files: make openapi regenerates the committed schema byte for byte. There are no SQL changes.
• No live model run: execution is unaffected.
• Independent blind review of the full diff by a fresh Claude Code subagent (the user-approved replacement for GPT-6 Astra): no code or security defects. Its in-scope findings (stale docs and missing write-route tests) are fixed in the last three commits, which touch only docs and tests.

Deferred

Not part of this change and registered on the board:

• The Environment Files list still rejects unknown keys through its own parser.
• Files limit=abc versus limit=-1 codes are inconsistent.
• Error order and empty-limit= messages.
• Malformed non-UUID path IDs return 400 in Core but 404 officially.
• Metadata/name error fields.
• Skill sole-version deletion and number reuse.
• Session delete lifecycle.
• Whitespace input.
• Template network forms.

No full protocol compatibility is claimed.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}