Catch up references_v3_compendium.md with everything since mid-August - #256

The compendium had not been updated since roughly the point
ResultsReferenceFinder3 and string interpolation shipped -- everything
built since then (a large fraction of the whole references v3 effort)
was undocumented here. This is a pure documentation update, no code
changes.

Added, in §5, in the same "settled facts, not chat narrative" voice
the rest of the document already uses:

- ReferenceResult3's actual four fields (path/uuid/data/identity --
  the existing description omitted identity entirely) and a table of
  what `path` actually holds per producer, since there is no
  discriminator field recording that.
- Rule 1 / Rule 1a / Rule 1b -- whole-resource content resolving to
  exactly one entity, and the two global-ledger exceptions. Rule 1b
  existed only in code comments before this, never in
  manifest_field_functions_proposal.md alongside Rule 1/1a.
- _check_position() -- the enforced, declarative replacement for
  scattered hand-written "is this recognized" guards, and the bug
  class it closed project-wide.
- The manifest field-accessor catalog (Part A/B) -- 34 field-accessor
  functions and :path() (Rule 2/3), the shared SOURCE/KEY/POSITIONS
  shape, without reproducing the exhaustive per-function table that
  already lives in the proposal doc.
- :having() and :from()/:to() ranges, at both version/run level
  (two modes: index and date) and statement level (index-mode only).
- ReferenceExpression3 -- UNION/SUBTRACT/INTERSECT semantics, the
  join-key rule, and the left-side-dedup bug caught by testing against
  a concrete worked example rather than design review.
- The full '*' traversal generalization across all three datatypes:
  field accessors, :having()/:flatten(), the :all() meaning-collision
  (resolved via a worked example, composite-key partitioning), path
  narrowing, name_three, and pointer optionality -- including the two
  previously-latent bugs each pointer-optionality fix surfaced, and
  the CSVPATHS-vs-RESULTS/FILES inconsistency found and deliberately
  left as a documented, not-silently-fixed gap.

Updated §6 (Known gaps) to mark the INTERSECT "later phase" note
resolved (ReferenceExpression3 is built) and add the one traversal
restriction that has survived every generalization pass:
:manifest() combined with real '*'-traversal narrowing, with the
specific result.path-vs-result.uuid disambiguation problem that blocks
it, planned as a dedicated next branch.

Rewrote §7 (Not yet built at all) -- nearly everything it tracked is
now built; collapsed those into a short pointer back to §5 rather than
repeating the design reasoning, and corrected the remaining genuinely-
still-true items (type-ahead, {...} interpolation evaluation, v3's
still-pending production integration) where they had gone stale
(e.g. "no VALUE-role function is registered today" -- now false, the
whole field-accessor catalog is VALUE-role).

Re-verified test counts while at it: tests/references/ 486 -> 1099
passed; full suite 2066 -> 2687 passed, same known 11-failure
SFTP/S3/Nos baseline throughout, unrelated to references v3.

Also updated the document's own header (stale references_notes/ paths
-> specs/references_v3/, and the "as of 2026-07-31" date) to reflect
the 2026-08-19 reorganization into specs/references_v3/ and this catch-
up pass itself.