docs(mdcode): restructure section 7 of the actions guide for narrative

[libei]

Section 7 of the actions guide, restructured for narrative. No new claims about
the runtime: this is reordering, cutting, and two accuracy repairs made while
compressing.

What was wrong

The section was ordered by machinery. Resolve/bind/apply, then how rows are
found, then guards. The payoff of the whole guide — the constraint declared in
section 2 finally stopping a write — arrived 150 lines in, behind a six-name
list of which field names count as identifying and a paragraph on type-driven
predicate dropping. ### Why a guarded action is refused, the actual message,
got 26 lines; ### Which rows a call touches got 71.

The two judge subsections were 58% of the section and carried nine bold notes
between them, each a true fact and collectively a wall.

What it is now

One question: what happens when somebody tries the write. Three kinds of rule
in escalating order, each shown allowing or stopping a real call.

subsection	what it answers
### What a run does	resolve, bind, apply — figure 3
### When a rule stops the call	a non-advisory expression guard refuses
### When the rule is a sentence	--judge, the verdict, the reason
### When the judge needs a fact	--judge-reads-store
### Which rows a call touches	demoted to reference at the end

Table: the new order of section 7 and the question each subsection answers.

The opening is untouched — it was settled in #434.

394 → 340 lines, 2250 → 1757 prose words. 129 of the remaining lines are fenced
blocks, which is the floor.

Reference detail cut to what changes something a reader writes

• The six identifying field names (name, full_name, fullname, title,
  label, display_name, case-insensitive, first-declared wins) become one
  sentence naming the pattern. Accepted cost: somebody whose field is called
  label can no longer confirm from the guide that it counts.
• The judge's read caps stop being quotable numbers — "a small budget of reads,
  each capped in rows and in the size of a value" rather than 4 / 20 / 200.
• The SELECT * FROM (...) AS judge_read LIMIT 21 wrap goes. What stays is the
  consequence: a write can still land through a function on PostgreSQL, and
  nothing confines a read to the tables your model binds.
• One fenced block is dropped, the resolution SQL, because figure 3 already
  shows it.

Paragraphs that argued rather than informed

Three, all the class #434 was cut back for:

• the defence of refusing over running unchecked ("Quietly ignoring a rule your
  model declares would be worse than no runtime at all…") → one clause;
• "Nothing in kcmd is an approver, and a refusal that left this out would read
  as the end of the matter" → dropped, the quoted escalate sentence speaks for
  itself;
• "because one rule showing all three branches is easier to follow than three
  rules showing one each" → "so one rule shows all three branches".

Also dropped: the paragraph explaining why the demo names no expression in
guards, which section 2 already covers, and the sentence defending figure 3's
shorter lookup.

The judge subsections now carry seven bold notes where they carried nine.

Two repairs found while compressing

**Keep the rule settleable from a few rows.** briefly read "A judge that still
can't tell answers that the rule doesn't hold". The source instructs the judge
to answer that way; it does not guarantee it. Restored to "is instructed to
answer".

**The rule is settled before the transaction opens.** ended "A rule about
either belongs in an expression", where either pointed at a state and a
race. Both are now named.

Claims checked against the source

claim	source
"Only a constraint the action names in guards has a say in a call" — broadened from "can refuse a call like this"	run_action.ts:37-38: "a rule takes effect where something references it, and guards is that reference for an action"
"a call that a non-advisory expression guards"	run_action.ts:492 filters on_violation: warn out of the guard set before the refusal test
"comparing the input against the column your entity's primary_key binds" (singular)	run_action.ts:995: the key predicate is added only when keyColumns.length === 1
"A name is also the only way to reach an entity keyed on several columns"	run_action.ts:990-991 and the composite-key error at :1012-1017

Table: each compressed claim and the source that settles it.

Three headings renamed, six inbound links updated

old anchor	new anchor	sites
#a-guard-settled-in-words	#when-the-rule-is-a-sentence	actions.md 414 and 1388, demo/semantic-model/agent/README.md 215
#a-guard-that-reads-a-row	#when-the-judge-needs-a-fact	actions.md 461, reference.md 89
#why-a-guarded-action-is-refused	#when-a-rule-stops-the-call	actions.md 1371

Table: the renamed anchors and every link that points at them.

The demo README link was outside the docs/ tree my first grep covered, and a
cross-file checker found it. Every anchor link under toolbox/mdcode now
resolves — 204 links across 32 pages — and a repository-wide grep for the old
anchors returns nothing.

#7-run-it is unchanged, so the four links to it and reference.md:80 are
untouched. The in-section self-reference at the end of ### When the judge needs a fact became "the race described above", since the two subsections are now
adjacent.

Fenced blocks

Blocks move, and one is dropped. Every other fenced line is byte-identical to
fa683d6:

-```sql
-SELECT account_id FROM account
-WHERE account_id = @ref0 OR name = @ref LIMIT 2
-```

84 fence markers → 82.

Checks

Prose wrap ≤ 80: clean. Internal anchors: 36 headings, 0 broken links. Banned
constructions: 0. "we": 0. Sentences over 40 words in section 7: 4 → 1, and
that one is the checker reading a lowercase kcmd as mid-sentence rather than
as the start of a new one.

Rebased onto fa683d6 (#433, section 8). Two conflicts, both where #433 revised
the same §8 sentences my anchor swap touched; #433's prose was kept and only the
anchor changed.

Interaction with #436 — one link, whichever merges second

#436 restructures section 8 and is branched off the same fa683d6. Merging both
leaves one dead anchor, and git reports no conflict at the site:

MISSING ANCHOR docs/semantic-model/actions.md -> actions.md#why-a-guarded-action-is-refused
32 pages, 204 anchor links, 1 broken

The sentence [refusal from section 7](#why-a-guarded-action-is-refused) arrives here instead sits at fa683d6 line 1443 and at 07a54bc line 1361 — #436
relocated it within section 8, so git can't match this PR's rename to it and
keeps #436's copy with the old anchor. The one conflict git does report is
elsewhere and resolves by taking #436's side.

Every other inbound link survives: the sites at 07a54bc actions.md 414, 461,
1263, 1497, reference.md 89 and demo/semantic-model/agent/README.md 215 are
unmoved from main, so the rename applies automatically. The reverse direction is
clean — nothing here points at the section 8 anchors #436 renames.

Whichever lands second updates that one anchor to #when-a-rule-stops-the-call.
Reported to the session that owns #436.

Still true after this PR

Section 7 says "Nothing in kcmd evaluates an expression against live data
today". #421 falsifies that sentence when it lands.

The action list block still shows payments/TransferFunds and was built
from the guide's own fenced YAML rather than captured from a live run. Worth
diffing against one; not in this PR.