refactor(mcp): _INSTRUCTIONS orients the workflow, not a rulebook (#4389)

Spike #4389 read the spec, Claude's docs and a dozen servers: the field is
for how the tools fit together, and the field runs ~600-1,600 characters.
Ours sat at the 2,048 cap as a keyword index that also carried stance.

- JUDGE, REPORT and MISSED leave the index. They fire mid-work, not at
  session start; using-scribe and reporting-back state them in full, and
  `placement`/`report_back` cue reporting in-band. No skill text changes.
- The rest is rewritten as plain practices (1,503 chars) and keeps every
  session-start marker the ownership registry pins.
- INSTRUCTIONS_BUDGET 2000 -> 1600; the three index markers are dropped
  from the registry; the miss-route index test now checks that the index
  keeps what_might_apply and stays off the route.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/test_guidance_ownership.py

@@ -25,7 +25,10 @@ WHAT IT PINS
    moments declares that in `shared_with`, with the reason beside it.
 3. THE INDEX NAMES THE SESSION-START REFLEXES. `_INSTRUCTIONS` is the one
    surface every MCP client receives, so each reflex it indexes keeps its
-   `index` markers there — a one-line pointer, not a copy.
+   `index` markers there — a one-line pointer, not a copy. Only session-start
+   practices carry `index`: judging, reporting and the missed-rule route fire
+   mid-work, and #4389 took them out of the index (the skills and the in-band
+   responses carry them at the moment they apply).
 
 WHAT IT CANNOT SEE
 
@@ -174,8 +177,7 @@ TOPICS: tuple[Topic, ...] = (
     # returned zero events. Machinery with no route to it.
     Topic("a missed rule is a trigger to fix before a floor to move", U,
           ("retrieval_telemetry", "tune_retrieval", "retrieval_surfaces"),
-          "take it to the record first and the dial second",
-          index=("retrieval_telemetry",)),
+          "take it to the record first and the dial second"),
     Topic("ask what already covers a moment before writing a record", U,
           ("what_might_apply",), "ask what already covers that moment"),
     Topic("reference notes update in place; dev-logs don't", U, ("reference note",),
@@ -193,7 +195,7 @@ TOPICS: tuple[Topic, ...] = (
           ("create_snippet", "when_to_use", "first build", "second copy"),
           "prior art offered beside a write is not noise", index=("create_snippet",)),
     Topic("report back where the work stands", "skill:reporting-back", ("reporting-back", "placement"),
-          "take the placement from the record", index=("placement",)),
+          "take the placement from the record"),
     Topic("the operator's own reply shapes come first", "skill:reporting-back",
           ("reply_preferences", 'content_type="rule"'),
           "the operator's own shapes come first"),
@@ -225,8 +227,7 @@ TOPICS: tuple[Topic, ...] = (
     # that trimming one cannot quietly take the other with it.
     Topic("the agent judges the record, and judging is attended", U,
           ("judge of record", "attended", "classify_shapes"),
-          "a record that is wrong stays wrong until something rewrites it",
-          index=("judge of record",)),
+          "a record that is wrong stays wrong until something rewrites it"),
     Topic("a finding surfaced and not judged is a finding dropped",
           "skill:reporting-back", ("you are the judge", "hard to reverse"),
           "it reads as diligence and functions as a backlog"),

tests/test_instruction_surfaces_agree.py

@@ -69,8 +69,11 @@ def _instructions_text() -> str:
 # Claude Code injects only the first ~2,048 characters of an MCP server's
 # instructions and silently cuts the rest mid-word (#2562: observed live —
 # the previous 20k-char version delivered ~10% of itself, and none of the
-# Systems tagging guidance ever reached a session). 2,000 leaves margin.
-INSTRUCTIONS_BUDGET = 2000
+# Systems tagging guidance ever reached a session). The budget is set below
+# the cap, not at it: the field runs ~600–1,600 characters (spike #4389), and
+# a block pressed against 2,048 is being used for depth that belongs in a
+# skill.
+INSTRUCTIONS_BUDGET = 1600
 
 
 def test_instructions_fit_the_fold():
@@ -181,9 +184,10 @@ def test_the_strength_guard_can_fail():
 # ── Reporting back (milestone 409 step 3) ──────────────────────────────
 #
 # The reporting-back skill carries the shapes, but a skill only helps if it
-# fires. The reflex that sends a session to it is stated in using-scribe; the
-# server's index carries a REPORT line pointing at `placement`, and the in-band
-# cue on update_task is the half that fires on its own in every client. The
+# fires. The reflex that sends a session to it is stated in using-scribe, and
+# the in-band cue on update_task (`placement`, `report_back`) is the half that
+# fires on its own in every client. The server's index carried a REPORT line
+# until #4389: reporting is a mid-work stance, not a session-start practice. The
 # static context was a second plugin-side copy until milestone 410.
 
 REPORT_REFLEX = "report back in a shape the operator can read"

tests/test_retrieval_miss_route.py

@@ -139,20 +139,19 @@ def test_writing_a_new_record_asks_what_already_covers_the_moment():
     assert "no bar" in text
 
 
-def test_the_index_points_at_the_route_without_restating_it():
+def test_the_route_stays_off_the_index():
     """`_INSTRUCTIONS` is the only surface every MCP client gets (decision #4027).
 
-    It indexes; using-scribe states. A client with no Agent Skills support still
-    learns the ordering exists and which tool opens it.
+    It carried a one-line MISSED pointer to this route until #4389, which found
+    the index was being used as a rulebook: a missed rule is noticed mid-work,
+    not at session start, and the spec asks server instructions for the
+    workflow across tools, not stance. using-scribe states the route (the tests
+    above pin it); the index keeps what_might_apply, the tool that opens it.
 
-    Whitespace-flattened before matching. The block is hard-wrapped to fit a
-    2,000-character budget, so a phrase straddles a line break the moment
-    anything before it changes length — #4103 shipped a guard that broke
-    exactly that way, on text nobody had touched.
+    Whitespace-flattened before matching: the block is hard-wrapped, so a
+    phrase straddles a line break the moment anything before it changes length
+    (#4103).
     """
-    text = " ".join(_instructions().split())
-    assert "retrieval_telemetry" in text
-    assert "not a floor to move" in text
-    # The route itself must NOT be here — 2,000 characters is the whole budget
-    # and Claude Code cuts the rest mid-word (#2562).
-    assert "take it to the record first" not in text.lower()
+    text = " ".join(_instructions().split()).lower()
+    assert "what_might_apply" in text
+    assert "take it to the record first" not in text