docs(semantic-model): restructure section 8 around its argument

[libei]

Follow-up to #433, which revised section 8's prose. This one changes its
structure and nothing else about the argument.

Replaces #436, which is the same work on a pre-rebase branch. #435 landing moved
main to 985c75b and left #436 conflicting; this PR is that work rebased, plus
a writing pass on top.

What was wrong

Section 8 ran 356 lines: a 152-line unheaded preamble (43% of the section)
followed by five H3s. Three consequences:

• A reader met [NOT RUNNABLE] in the listing at line 1301, learned who adds
  it at 1412, and learned why at 1441 — a 215-line gap, for the marker
  that carries the section's point about binding.
• Reference material was 22% of the section, spread across five non-adjacent
  places.
• Calling it from code ran 74 lines over three distinct topics.

The underlying cause is that the section does two jobs in one stream: it makes
an argument (you don't write these tools; they're derived) and it is the
reference for the derivation. Those want opposite orders.

What changed

Six headings, each with one job; the preamble drops to 15 lines because the
derivation and its three products are named up front instead of 143 lines in.

heading	job
## 8. Hand it to an agent + 15-line opening	the claim
### The set an agent is handed	command, listing, and the [NOT RUNNABLE] gloss in place
### Where each line comes from	Table 4, and who owns each half of the instruction
### What a write tool and a lookup tool do	unchanged
### What a withheld tool is waiting on	the reasons, with the --judge contrast beside the reason it demonstrates
### Calling it from code	unchanged
### The commerce demo, worked through	unchanged

### A tool says whether it can be called is gone; its opening paragraph moved
up under the listing and the rest became What a withheld tool is waiting on.
### Who owns the instruction is folded into the provenance heading, where
Table 4 has just shown ai_context.instructions becoming the instruction.

Six exceeds the guide's usual five H3s per section. Five was reachable only by
merging tool provenance with tool capability, which would recreate the
74-line two-topic heading this is meant to fix.

One claim corrected

kcmd agent tools did not "read your model and nothing else". It resolves the
store your profile binds (commands.ts:1352, :1379), and a model whose
profile binds none prints offers no tools under this profile. and exits
non-zero (:1387). The same sentence also said it opens no session and changes
nothing, which is true — runtime.ts:176 resolves a store without
connecting, and the listing calls no model.

Writing pass on the restructure

Six prose defects, in the second commit:

• A garden path in subject position — "A tool this binding can't serve is…"
  reads "A tool this binding" as one noun phrase.
• withheld first appeared in a heading, 100 lines after the concept it names.
  The section was calling one thing by three names ([NOT RUNNABLE], "tools
  this binding can't serve", withheld); they are bound together at first
  contact now.
• What a withheld tool is waiting on opened on two library field names before
  the reader has a reason to care. Those moved to Calling it from code, which
  owns the library surface.
• "derived from it" — the nearest noun is "the change", not the model.
• A third "It" in a chain whose previous link is the store, which can itself
  open a connection.
• "offerable" where the document already says callable.

Anchors, after #435

Section 7 renamed three headings. Git remapped six of the seven references
automatically during the rebase, but not the one this branch had relocated
within section 8 — git only maps a rename onto a line that stayed put, so that
site came through with a dead target and no conflict marker. Fixed explicitly:

#why-a-guarded-action-is-refused  ->  #when-a-rule-stops-the-call

The new target is a closer fit than the old one: ### When a rule stops the call is the subsection carrying the TransferFunds / AmountIsPositive
expression refusal this sentence cites, and it ends on "a refusal is settled
before a session opens, so a refused action leaves no transaction behind" —
which is the claim the section 8 sentence makes.

Verification

• All five fenced blocks carried across byte-identical (multiset diff
  against origin/main: 0 lines removed, 0 added).
• Prose at 80 columns; every internal link resolves; frozen anchors intact.
• Old anchor targets repo-wide: none. Stale link label text: none.
• xlinks.py over the tree, node_modules excluded: 0 broken.
• bun test: 1083 pass, 0 fail.

Net: 356 → 354 lines.