feat(excel): formula dependency trace — get --prop trace=precedents|dependents

[ylz92871-dotcom]

feat(excel): formula dependency trace — get --prop trace=precedents|dependents

What: get <file> '/Sheet1/F2' --prop trace=precedents [--prop depth=N]
returns the formula dependency graph around a cell (dependents likewise).
Precedents collect references from the evaluator's token stream — there is
no table-formula AST, and defined names are already folded by the tokenizer,
so cell, range, sheet-qualified, and named references all come out of one
pass. Range leaves expand with caps (<=5000 per range, 20000 leaves / 5000
edges overall; overflow sets truncated: true). Dependents answer from a
lazily-built in-process reverse index (every formula cell scanned once,
range refs clamped to the used range), invalidated on every mutation: the
row-index hook, the Set() entry, sheet rename and sheet remove. JSON adds
format.trace = {root, edges: [{from, to[]}], truncated} to the existing
cell node — nothing else in the envelope changes; text mode renders an
indented tree. Errors are explicit with codes: precedents on a constant ->
unsupported_element (+suggestion), missing cell -> not_found, bad depth or
unknown prop -> invalid_argument.

Why: formula audits today stop at the cell — nothing shows what a formula
depends on or what breaks when an input changes, so agents re-derive
dependency chains by grepping formula text. A token-stream trace is exact
for the grammar the evaluator itself accepts (including the same defined-
name folding), needs no evaluation, and gives the dependency question a
one-command answer.

How to verify:
officecli get report.xlsx '/Sheet1/F2' --prop trace=precedents --json
officecli get report.xlsx '/Sheet1/F2' --prop trace=precedents --prop depth=3 --json
officecli get report.xlsx '/Sheet1/B2' --prop trace=dependents --json
(drop --json for the indented text tree)
Build 0 errors; local harness T-11 21 assertions green (chain + cross-sheet
/Sheet1!B1 qualification + reverse-index invalidation rebuild + range-entry
matching + truncation caps + text tree); regression moat 26/26.