Return official errors for unresolved list cursors

[SaladDay]

When a list's after cursor does not resolve, each list now returns the officially observed error for its family. The pinned baseline is unchanged (SDK 3.13.0 / d7c41ef / agents=v1).

Behavior

Lists	Unresolved cursor (random, malformed, other type, other parent, deleted, foreign)
Agents, Sessions, Turns, Templates, Vaults, Credentials	404 not_found_error, param null. A malformed cursor now takes the missing-ID path (it used to be 400 invalid_request)
Session Items, Subagent Items, Subagent Turn Items	400 invalid_request_error, param null, Invalid session item ID in `after`
Subagents, Subagent Turns	400 invalid_request_error, Invalid resource ID in `after`
Artifacts	400 invalid_request_error, after is not a valid artifact ID
Skill versions	Not a version ID: 400 invalid_value, param after. Missing: 404. Another Skill's version (same tenant only): 400 invalid_value, "Skill version cursor does not match this skill."
Files, Skills, Environment Files	Unchanged; they already matched

• Parent check first: a missing or foreign parent is still checked first and returns 404. A foreign cursor is byte-identical to a missing one. Artifact and Skill-version lists re-check the parent before returning a 400, so a parent deleted mid-request still gets 404.
• Bounded echo: the Skill version and Skills order errors echo the caller's value only when it is at most 256 bytes of printable, valid UTF-8 (shared internal/echotext). Otherwise the message omits the value.
• Unchanged: valid-cursor paging, ordering and limits. The Core-only Runtime observations list shares the Session cursor and now answers a malformed cursor with 404 (documented).
• Upstream anomalies not copied: the official 500s seen for some cursors. Deleted Agent/Session IDs that still work as official cursors (ERR-15) would need tombstones and are deferred.

Evidence

• Campaign scan 4 (105 owned official requests, 2 Sessions, 4 Turns, all deleted): ERR-01..06.
• SAT-04 (scan 3) and HE-57 (scan 2).
• Recorded in list-query-semantics.md and operation-evidence.md (register CE and the list rows).

Validation

• Independent acceptance (written from the requirements only; real Core with PostgreSQL; raw HTTP and the pinned SDK; 968 requests with seeded Subagent/Artifact rows and a no-write fingerprint):
  • On main, the 286 changed checks failed and the 1,002 retained, isolation and paging checks passed.
  • On the final head, all rows pass. Foreign cursors are byte-identical to missing ones.
• Go tests: a real-PostgreSQL public test compares exact bytes for every row for two tenants and pages every list with has_more/first_id/last_id checks. Echo-bound tests cover 257-byte, 200 KB, control-character and invalid-UTF-8 values. The pinned-SDK list-query script passes.
• Server gate on this head: all make check targets plus the Web typecheck, core-doctor, unit tests and build pass. The Playwright browser cases were not run on the server (no Google Chrome; the user approved the skip). make sqlc-generate and make openapi are byte-identical.
• Independent blind review by a fresh Claude Code subagent (the user-approved replacement for GPT-6 Astra) found no tenant or parent leak. Its blocker, the unbounded echo, and its follow-ups (docs, stronger paging checks, the Artifact race) are fixed in the last commits. These were verified by the unchanged acceptance script and the gate, without a new review, per the user's rule.

No full protocol compatibility is claimed.

---

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