…es them

Inverts the origin-description rule: parameter descriptions must not contain a
registered tool name (that coupling goes stale on rename and was already wrong —
three descriptions named list-available-fields as the producer of a path it
never returns). Descriptions now state the semantic source and, for optional
opaque params, what happens when omitted. The gate enforces this surface-wide.

Error text keeps naming tools — the reachable-tools gate makes that safe — so
the apply-* recovery messages now name get-worksheet-xml / read-cached-xml /
write-cached-xml instead of periphrasis.

Also: resolve-field's stale-cache guard no longer requires session (the
description tells agents to omit it), and the byte-budget grandfather caps for
add-field and inject-template ratchet down to their new smaller sizes.

Co-Authored-By: Claude <noreply@anthropic.com>

…out env var

The apply seams resolve an entity-encoded or NFD sheet name to its canonical
form but forwarded the caller's raw name to activation, so navigation silently
dropped. The seam now overrides focus.sheetName with the canonical name — fixed
once, not at nineteen call sites.

TABLEAU_DESKTOP_CALL_TIMEOUT_MS above 60s did nothing: production never passed
clientOptions. The session manager now wires config.desktopCallTimeoutMs into
the External API client.

activate-sheet's not-found error carried a bare structuredContent block with no
message and no next action; it now uses the same envelope as every other
producer, and a new source-scan test enforces that invariant surface-wide.

Co-Authored-By: Claude <noreply@anthropic.com>

…r has

Six live strings still told the agent to call apply-* with mode=file — a
parameter DEC 7 deleted and the served schemas now reject. They name the cache
path contract instead. The unreachable over-cap note and its green test are
deleted. One dialog warning spelled execute-tableau-command with underscores;
fixed, and the reachable-tools gate now catches underscore spellings too. The
search recommendation no longer ends at an XML route the default profile
cannot take.

Knowledge section reads no longer throw URIError on headings containing a bare
percent sign (a real shipped heading has one); the raw fragment is the fallback.

The field-reference regex anchors the role, so a five-segment column instance
can no longer silently rename a field to its derivation suffix. The corpus test
pins the eleven cue outcomes skips were hiding and renames five tests to claim
only what they prove.

Co-Authored-By: Claude <noreply@anthropic.com>

…iter; 2.57.2

The consumer repo (agent-to-tableau-desktop) must re-sync these two
lockstep-core files; logged on the port-debt ledger.

Co-Authored-By: Claude <noreply@anthropic.com>

… manual

Co-Authored-By: Claude <noreply@anthropic.com>

…he mocks

The re-review of the fix round confirmed four P1s and two P2s; all fixed:

- The reuse anchor is now captured AFTER the apply. It was stored pre-apply, so
  on live Desktop (where the apply itself advances the event sequence) every
  repeat looked user-edited and the dedupe would never have fired — the mocks
  hid it because their events never advance.
- Reuse that authored calcs reports applied:true; the workbook was written.
- The recovery gate admits any call naming target_worksheet — an explicit
  rebuild instruction is not a bare resubmit — and reuse no longer erases
  recovery state, so the repeat-bind breaker still engages on bare loops.
- The multi-encoding steer chains worksheetFile through consecutive add-field
  calls instead of restarting each from the live sheet.
- columnRef and fieldMapping descriptions state their provenance again (from
  field resolution, never invented) without naming tools; byte caps ratchet
  down to the new smaller sizes.
- The surface-wide session-never-required assertion is restored across the
  full registry.

Still unproven live (named in the PR): the post-apply anchor against a real
Desktop event stream; transports without getEvents store no anchor and the
receipt's presence-only caveat covers them.

Co-Authored-By: Claude <noreply@anthropic.com>