docs(adjudication): decide how anchors and transclusions resolve

[CameronBrooks11]

Adds docs/adjudications/0003-anchors.md, which decides how comments, suggestions and transclusions anchor to text in a blockspec document. The aim is that after edits and a stock git merge, an anchor either lands on the right text or is refused loudly. It also adds SPEC.md §11, which carries the resulting rules.

Stacked on #40, whose SPEC.md this extends. The base is docs/spec-draft-0, so the diff shows only this change.

No implementation and no conformance cases. The reference resolver and the cases come later, and the cases come first, from an author who has not read a resolver.

What 0003 decides

	decided	where
Q1	one record file per comment or suggestion, in <doc>.notes/	0003 §2, SPEC §11.1
Q2	comments follow a moved leaf, marked MOVED; suggestions never apply to one	0003 §5–§6, SPEC §11.6–§11.7
Q3	an unreachable base is REFUSED(base-unavailable). Before landing, a tool MAY re-base each record that resolves onto the branch head. A CI check MAY require that; records that do not resolve keep their base, are reported, and do not fail the landing	0003 §1, SPEC §11.11
Q4	a transclusion is refused once its name is rebound, until the referrer re-pins	0003 §7, SPEC §11.9
Q5	transclusion is a titled link, [t](doc.md#name "transclude"), amending 0001	0003 §7, SPEC §4.2, §11.9
Q6	the refill rule is kept	0003 §4, SPEC §11.4
Q7	resolving through merge parents is a tool's option that only adds refusals, outside the format (Q9's identity tracking likewise; Q3's re-basing is not a refusal rule)	0003 §1, §9, SPEC §11.11
Q8	records are not deleted in normal use; a purge removes a whole thread; an orphaned reply is #REF!(reply-to)	0003 §2, SPEC §11.8
Q9	the notes directory is authoritative; a rename moves it in the same commit; a duplicate record id across directories is warned on; a tool MAY track document identity	0003 §2, SPEC §11.8, §11.11
Q10	only status: open suggestions apply	0003 §6, SPEC §11.7
Q11	write-back only through an exact NFC-to-raw mapping, else unmappable-offset	0003 §6, SPEC §11.7
Q12	the twin narrowing: only an unmatched exact copy refuses an edited leaf	0003 §4, SPEC §11.4
Q13	32-code-point context, no comment-only fallback, until a real-history replay	0003 §5, SPEC §11.5

0003 also fixes these details beyond the thirteen answers, and lists them under its decision table:

• the resolution reasons and their names, including not-in-profile for any snapshot the profile refuses, and #REF!(name);
• record-inconsistent for a short context away from the leaf's edge, for an empty span, and for a pinned name that binds nothing;
• replace is written exactly as stored, never normalised.

It also covers:

• the resolver's steps: correspondence (§4), then the span (§5);
• the resolution reasons, listed in SPEC §11.10;
• a table of the outcome each edit shape must produce (§8);
• the residual (§9): four mechanisms whose bytes match a legitimate edit. R1 is a deletion plus a verbatim copy elsewhere. R2 is a rewrite beyond recognition with the old wording left in order. R3 is a sentence rewritten inside a leaf with the old one re-added in the same context. R4 is a different document at the same path. The section says which of them reach suggestions and what would close each.

spike/LOG.md §9 is cited as the published single-commit instance of that family.

No figure in 0003 measures these rules on real history. Its What would reverse it starts with that replay.

Outline

0003:

1. what the format guarantees, and what it cannot
2. where anchors live
3. the record
4. which leaf: correspondence and when to refuse
5. which characters: the span, and statuses
6. suggestions
7. transclusion
8. what each edit must produce
9. the residual
10. prior art
11. the cost of reversing each choice
12. what would reverse it

Then the effect on 0001 and 0002, and what draft 0 carries.

SPEC.md §11 has twelve subsections:

1. records
2. leaves, keys and similarity
3. correspondence
4. the anchored leaf
5. the span
6. statuses
7. suggestions
8. replies, deletion and document identity
9. transclusion
10. resolution reasons
11. what a tool MAY add
12. what this does not prevent

The former §11, Not decided, is now §12.

Also edited:

• SPEC.md §4.2 names the one inline shape with meaning, and §8.3 says slugs are not transclusion targets
• README.md and AGENTS.md list 0003 and say it has no cases yet
• DESIGN-BRIEF.md §2 points its quote-anchoring question at 0003

Open questions 0003 leaves undecided

These are listed in SPEC.md §12:

• the exact syntax of a record's frontmatter: which YAML, quoting, and where reply-to sits
• how a referrer's pins are written in its frontmatter
• a table leaf's leaf.text, since §6.3 gives a table's content as rows of cells
• which chain to take when several are longest, which decides EXACT against MOVED
• whether the 0.5 and 0.15 thresholds are compared exactly or in floating point
• whether a hit must reach the leaf's edge when the stored context is short because the edge was nearer
• transclusion details: how path resolves, a destination with no path, the title's other quoting forms, a transclude link among other text, a transclusion inside a container, a transclusion with no pin, and self-transclusion
• what a validator reports for a rename that leaves its notes directory behind and for a duplicate record id, and whether either has an identifier
• the thresholds themselves, which come from D8 and are untested for this resolver
• range anchors, a selection across leaves, still open from 0002 §2

Checks

The case tree and the reference reader are unchanged. Nothing implements 0003 yet.

$ just check
All checks passed!
$ just conform
0 failure(s) across 237 case(s) in the fixture tree
$ just tree
0 problem(s)
$ just test
8 passed in 1.57s