feat(desktop): exact column_ref first-class + shared ref helpers (a2td #217+#222 port, seam-1 packet A)

[mattcfilbert]

Port-wave packet A (a2td #217 + #222 twins), from the content-signature port audit — the resolver half of the seam-1 multi-datasource correctness cluster, worth having in the feature/desktop binary: a duplicate Profit across datasources currently resolves by name-affinity and can silently bind the wrong one.

#222 half: shared exported ref helpers in field-resolver.ts (COLUMN_REF_REGEX, parseDatasourceQualifiedColumnRef, parseColumnInstanceRef, parseCanonicalColumnRef, formatCanonicalColumnRef — same API as a2td); private regexes in fields.ts/explicit-bind.ts replaced. Documented divergence preserved (explicit-bind accepts bare [deriv:Base:sfx]; fields require datasource-qualified). Colon-tolerant instance parsing — the minion's red run caught [Profit:Ratio] being mis-parsed as [Profit] and pinned it.

#217 half: resolveField and proposal-binding resolveInSummary recognize an exact canonical column_ref FIRST — a hit resolves exact; a ref-SHAPED miss fails closed as not_found (never falls through to fuzzy). Datasource selector matches internal name or UNIQUE caption; ambiguous captions refuse with guidance (Parameters excluded). Gate 5b single-datasource closure untouched. listAvailableFields untouched (#535's file — coordinated, not collided).

Validation (orchestrator, firsthand): full suite 4,362 green (298 files), all three wrong-bind invariant suites green (210 tests — portability, bind-behavior-matrix, carrier-uniqueness), tsc clean, lockstep 6/6, eslint clean (4 fixable errors fixed at adjudication). Version 2.25.4 assumes #538 (2.25.3) merges first — resolve upward on conflict.

Companions: #538 (receipt honesty, packet C) already open; packet B (planner/apply datasource selection, a2td #219+#220) stacks on this branch next.

🤖 Generated with Claude Code

---

Adversarial review adjudication (fresh-context GPT review, orchestrator-verified): the review returned two findings, both refuted by reference-parity — (1) ref-match-before-name ordering in resolveInSummary is byte-for-byte the a2td-merged design (explicit comment in a2td validate.ts:217: refs are already datasource-qualified, so they resolve before duplicate captions — the ordering that closed #217's bug class; the trigger requires a field named as another field's canonical ref string, adversarial not user input); (2) internal-name-beats-caption selector precedence is identical in a2td's matchingFieldsForDatasource — deterministic, with explicit refusal on caption-side ambiguity. Review's clean lenses: call-site enumeration, colon-parsing edge cases, old-vs-new regex grammar preservation. No port-introduced behavior deltas found.