Accept ARRAY vector literals and typed value errors

[EnRaiha]

Summary

ARRAY[0.1, 0.2, 0.3] folds its literals to Decimal in the planner. Strict element coercion kept Float and Integer, dropped every other element, and reported the literal as having 0 elements — and that mismatch surfaced as XX000, an internal error, for what is a client value problem. A non-numeric element was equally opaque: no element index, no type, and the same internal class.

This PR fixes the coercion and the class. Elements of every numeric shape are read; a bad element is refused by index and type; and a client value error keeps its BadRequest class across the Data Plane boundary instead of degrading to XX000.

Behaviour changes

Path	Before	After
INSERT … ARRAY[0.1, 0.2, 0.3] into VECTOR(3), strict schema	XX000, "got 0 elements"	the row inserts
ARRAY[0.1, 'nope', 0.3]	XX000	42601, naming the element index and its type
ARRAY[NaN] / ARRAY[Inf]	folded through	refused
A client value error crossing the Data Plane or the cluster wire	XX000	42601

Closes #306.

Root cause

• extract_vector_floats in strict_format/coerce.rs matched Value::Float and Value::Integer only. The planner's literal folding produces Decimal, so a literal ARRAY[0.1, 0.2, 0.3] contributed zero elements and the column reported got 0 elements.
• The follow-up problem was the error class, not the coercion: crate::Error::BadRequest had no mirror in the bridge's ErrorCode enum, so the conversion fell through the catch-all to ErrorCode::Internal { detail } — every client value error crossed the Data Plane boundary as XX000, even though the direct (single-node) path already assigned 42601.
• Two strict-encoder wrap sites (handlers/convert.rs, point/update/post_image.rs), the update-from-join collect pass, and two doors found by the audit rounds (below) had the same fold-into-Internal shape.

What changed

Commit 0fd44a626 — coercion:

• extract_vector_floats reads Float, Integer, Decimal, and numeric String elements; a non-numeric element is rejected with its index and type; non-finite values are refused with the value.
• The four binary_tuple wrap sites (stored_body.rs, stage_write/body.rs ×2, stage_upsert.rs) preserve BadRequest instead of wrapping it.

Commit ddefcaa0f — wire coverage:

• document_strict_inserts_a_decimal_array_literal_into_a_vector_column (the filed case)
• document_strict_reports_a_non_numeric_element_by_index_and_type (message names the element; SQLSTATE pinned)

Commit 6c29ac9fd — the bridge:

• ErrorCode::BadRequest { detail } added to the bridge envelope enum with the same doc-comment pattern as its siblings, and preserved in From<crate::Error>.
• Mirrored on the cluster wire (DataPlaneErrorCode, appended variant; the enum's own doc says variants are appended, never reordered) with both round-trip conversions.
• pgwire renders it 42601 (the code the direct path already assigns). The node-abort classifier treats it as not established: the same code can be raised after an earlier step of a multi-step plan applied, so refusal-without-apply is not provable from the code alone.

Fixes from the audit rounds:

• f912fb2b1: two more doors preserved BadRequest — handlers/convert.rs (CONVERT row encode) and point/update/post_image.rs (strict re-encode). Tests added: the abort classifier covers BadRequest; the cluster wire round-trips it verbatim.
• e3e129563: the update-from-join collect pass (handlers/update_from_join.rs), which encodes strict post-images, preserved it too.

Regression proof

On base main, without this change: the filed case reports expected VECTOR(3), got 0 elements (SQLSTATE XX000), and the bad-element case finds no element message. With this change both pass, and the bad-element SQLSTATE is asserted as 42601.

Tested

• cargo nextest run -p nodedb --test wire -E 'test(cases::engine_surface_document_strict)' — 2 new cases pass
• error-path regressions: 42 passed across error_from_data_plane, data_plane_error_wire, write_abort, and the sqlstate/error-code unit tests
• convert/typeguard modules: 22 passed (sql_convert_column_defs, sql_typeguard_default_gate, sql_typeguard_defaults)
• cargo fmt --all -- --check — clean
• cargo clippy -p nodedb -p nodedb-cluster --lib -- -D warnings — clean
• preflight — pass

Review

A separate, read-only parity audit ran over this branch before submission. It found the same defect class on further sites and asked for two tests; all are fixed and verified here.

Finding	Resolution
Two strict-encode doors folded BadRequest into Internal (handlers/convert.rs, point/update/post_image.rs)	both preserve the class; the abort classifier and the cluster wire round-trip are covered by tests (f912fb2b1)
The update-from-join collect pass folded it too	preserved on that path; targeted tests pass (e3e129563)
Cross-version compatibility of the appended wire variant	the enum's append-only discipline; CI covers the workspace build

Notes

• Host note: on this machine the native wire tests overflow the default thread stack and need RUST_MIN_STACK=16M. The same abort reproduces on an unrelated branch with no native-path changes, so it is a host property, not this change. CI runs them at the repository default.
• The wire variant is appended and never reordered; existing slots are untouched.

Commits

• 0fd44a626 — fix(strict-format): coerce ARRAY vector literals into a strict VECTOR
• ddefcaa0f — ARRAY vector literal wire coverage
• 6c29ac9fd — fix(bridge): carry a bad request across the Data Plane boundary
• f912fb2b1 — fix(executor): keep a bad request out of XX000 on the convert and post-image doors
• e3e129563 — fix(executor): keep a bad request typed on the update-from-join collect pass

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.