fix(member-fields): reconcile rows that carry no referenceName

[2000game]

Review follow-up to #159 (7 findings from /code-review high).

High

• A live row carrying no referenceName was matched by the documented name fallback and then rejected by every consumer (memberFieldReferenceName(row) !== spec.referenceName), so it could be neither updated nor created — the plan went INCOMPLETE and ct apply aborted the entire run. Only a row carrying a different reference name is a contradiction now (conflictingReferenceName), in the fold, in apply and in the resolver.
• The identity-mismatch diagnostics offered only ct destroy --member-field …, which is destructive and, for a row without a reference name, did not even match: destroy resolved the live row by local key alone, reported "already absent", dropped the state binding and let the next apply POST a duplicate. destroy (and its backup fetch) now prefer the state-bound id, and every message names the non-destructive fix — declare the live referenceName — beside the destructive one.

Medium

• Delete-candidate false positive: declared rows were tested against a set of declared reference names via ?? "", so a declared field whose live row has none was reported as "not declared — DELETE CANDIDATE". The pass now skips rows a declaration claimed.
• A ref into a group that is adopted but declares no memberFields fell back to the raw ref string with exact matching, hard-throwing on live stand-bewerbung vs. ref.groupMemberField(g, "stand_bewerbung"). With no declaration there is no exact identity to hold the row to, so those refs keep the normalised local-key match.
• A live field sharing only a declaration's display name (its own CT-minted referenceName) turned a create into a fold error and thereby aborted apply for every other resource. It now warns, names the referenceName to declare, and plans the create. A near-identity — a live reference name differing only in punctuation or case — is still refused.

Low

• The duplicate-match error no longer claims "no row has the exact referenceName" when the matches are precisely the rows that do.
• The resolver renders a missing reference name as missing and no longer asserts a state binding that may not exist.

Also: the actual side reports the declared reference name for a row that carries none — ct never PATCHes one, so diffing it would propose the same update forever.

Docs re-read and re-signed. npm test (1083 passing), tsc --noEmit, eslint src tests all clean; 9 regression tests added.

https://claude.ai/code/session_01PNmsG3pmvxQNA9sNmnAvYJ