fix(baseline): key fallow member/component findings by <parent>.<member>

[robmclarty-minga]

Problem

The dead fingerprint extractor keys a finding on <file>:<symbol>, where the
symbol comes from the first non-empty field in SYMBOL_FIELDS
(export_name, name, member, …). But fallow reports class/enum members
with parent_name + member_name, and component inputs/outputs with
component_name + input_name/output_name — none of which are in that list.
So every unused member of a file falls back to a path-only key and they all
collapse onto one fingerprint.

That collapse breaks gating, not just precision. fallowVerdict only masks a
slot green when findings.size === issueCount (the fullyTracked guard). A
member-heavy report fingerprints to far fewer keys than its issue count, so the
guard treats it as untracked and the baseline can never mask it green.

On a real 7,970-finding repo:

Extractor	Unique fingerprints	vs total_issues
before	3,148	−4,822
after	7,949	−21

Fix

• Compose a member identity <container>.<leaf> from
  parent_name.member_name / component_name.input_name /
  component_name.output_name, before falling back to the single-symbol
  fields. Members in a file are now distinct, e.g.
  dead-code:unused_class_members:src/a.ts:Svc.foo.
• Stop fingerprinting the workspace_diagnostics and next_steps arrays.
  Neither is counted in total_issues, and workspace_diagnostics items carry
  a path, so fingerprinting them pushed findings.size past issueCount —
  which the fullyTracked guard also (correctly) rejects, wrongly marking an
  otherwise fully-tracked slot untracked.

Scope / honest limitation

This does not guarantee fullyTracked for every large report. Distinct
findings that share a line-free identity — overloaded members, a symbol
re-exported or imported twice in one file — still legitimately collapse (the
residual −21 above). That's inherent to fingerprints that deliberately exclude
line/column. For grandfathering a large legacy dead slot, fallow's native
per-category regression baseline remains the right tool; this change makes the
fingerprint baseline correct and precise for the common member/component case
and lets member-level fixes ratchet individually.

Verification

• pnpm exec vitest run — 427 passing, incl. 3 new fallow cases
  (member composition, component composition, workspace_diagnostics skipped).
• tsc --build clean.
• Ran the built extractor against the real report above to confirm 3,148 → 7,949.
• node dist/cli.js (self-gate) green except the pack slot, which fails
  identically on pristine main in my environment (local pnpm version; a direct
  pnpm pack --dry-run exits 0) — unrelated to this change.