ARRAYFORMULA parity: broadcast-allowlist expansion + control-flow broadening (implicit intersection gated separately)

[hhimanshu]

Scope amended 2026-07-25

Original scope ("ARRAYFORMULA parity beyond basic spill, incl. implicit intersection") was post-v1/tracking-only per 2026-06-07-workbook-v1-scope.md Decision 2. That row is now superseded by 2026-07-25-amend-workbook-v1-scope-arrayformula-format.md (Decision 1), owner-approved via truecalc/distributions#64.

Approved, in scope now:

• Expand the per-cell scalar broadcast allowlist in arrayformula_lazy_fn (crates/core/src/eval/functions/array/mod.rs:1470) beyond LEN/UPPER to the full set of Sheets per-cell scalar functions (fixture-confirmed, not assumed).
• Broaden broadcast beyond IF to other control-flow-shaped functions (IFS, SWITCH), contingent on Sheets confirming per-cell behavior for each.

Listed but NOT yet cleared — requires its own explicit sign-off before implementation:

• General implicit intersection for unmarked range references in ordinary (non-ARRAYFORMULA) formulas (e.g. =A1:A10+1). This changes scalar-context evaluation semantics workbook-wide, not just inside ARRAYFORMULA — a materially bigger blast radius than the two approved bullets above. Do not start this without a separate go-ahead.

Still explicitly out of scope, unchanged:

• Structural-edit reference rewriting (insert/delete row/col) — remains a separate v1.1+ tracking issue (post-v1). Nothing in this issue implies or requires it.

Blocked on fixtures. Existing ARRAYFORMULA fixture coverage is thin and doesn't cover the broadened allowlist. New fixture rows (per newly-broadcast function) are needed via the truecalc/fixtures GAS pipeline before implementation lands — values are never self-confirmed. This is being sequenced together with other fixtures-needing issues; do not start core work ahead of that landing.

[hhimanshu]

Concrete examples for scoping "beyond basic spill" when this gets picked up:

1. Forcing scalar-only functions to run elementwise across a range:
   =ARRAYFORMULA(VLOOKUP(A1:A5, Sheet2!A:B, 2, 0)) — VLOOKUP normally takes one lookup value; wrapped in ARRAYFORMULA, Sheets runs it once per row in A1:A5 and spills 5 results.

2. Implicit intersection (no ARRAYFORMULA wrapper):
   B1: =A1:A10 + 1 — classic Sheets/Excel semantics implicitly intersect the range down to the row matching B1 and return a scalar, rather than erroring or spilling.

3. Logical functions broadcast across ranges:
   =ARRAYFORMULA(IF(A1:A10="", "", A1:A10*2)) — IF's condition isn't naturally array-vectorized; ARRAYFORMULA must broadcast it.

4. Concatenation broadcast:
   =ARRAYFORMULA(A1:A10 & " - " & B1:B10)

Still gated behind a superseding ADR per the tracking note on this issue — do not build without one.

[hhimanshu]

Plain-language summary of what's actually approved here:

Think of ARRAYFORMULA as a wrapper you put around a formula to say "run this once for every cell in this range." Two changes are approved:

1. Let more functions work inside that wrapper. Right now, only LEN and UPPER work correctly per-cell inside ARRAYFORMULA. This expands that to more text functions, IS* checks, rounding, dates, etc. Today, using those functions inside ARRAYFORMULA doesn't work right — so this only ever turns a broken thing into a working thing. It can't break anything that currently works, because it currently doesn't work.
2. Let IF-style logic work inside it too — extending the same per-cell trick to IFS/SWITCH.

Both are "safe" for the same reason: they only kick in when someone explicitly writes ARRAYFORMULA(...). Nobody's existing spreadsheet changes behavior, because nobody could have been relying on something that didn't work before.

Why this is blocked on fixtures: TrueCalc's whole pitch is "we calculate exactly like Google Sheets." The only trustworthy way to prove that is to type the formula into a real Google Sheet, see what Sheets actually returns, and save that as the correct answer (a "fixture"). The new functions above have no saved answers yet — nobody's recorded what e.g. ARRAYFORMULA(ISBLANK(...)) actually returns in real Sheets. Without that, there's no way to verify the new code is right, so implementation waits for that answer-key to be collected.

Not yet approved, listed separately above: "implicit intersection" for ordinary (non-ARRAYFORMULA) formulas — that one could silently change behavior of formulas that already exist today, so it needs its own explicit go-ahead before anyone touches it.