RFC-007 Phase 3: wiki-link + composes-with edges with slugify() resolution and dangling tracking (fixes #589)

[lantisprime]

Implements RFC-007 Phase 3. Closes #589.

What ships

Two rule-to-rule edge types on em-graph.mjs, both opt-in (registered in EDGE_TYPES, absent from DEFAULT_EDGES):

• wiki-link — [[name]] in rule bodies, fenced blocks and inline backticks excluded
• composes-with — top-level bullets under ## Composes with, nested bullets ignored

Plus the shared target resolution both depend on, per-edge-type dangling tracking, and a 31-case suite wired into CI.

The issue understated the work

#589 described two parsers and a fixture. Adding edges to rule nodes broke three assumptions Phase 2 had made explicit in its own comments:

Site	Before	After
--orphans	reported every rule unconditionally	reports a rule only at degree 0
--from <rule-slug>	early-returned a single node	falls through to the BFS, making good on the comment left at em-graph.mjs:290
--hubs	degree map was episode-only	rule ids seeded, so rules can rank

Every line anchor #589 cited was also stale by +10 to +25 lines, drifted by Phase 2 (#604). Several pointed at blank lines; two pointed at the wrong JSON object entirely. All corrected.

Target resolution needed a decision

Built literally, the spec resolves almost nothing. Measured against the live corpus before writing any code:

Edge type	Spec as written	With filename-alias index
wiki-link	13 of 132	122 of 132
composes-with	0 of 76	51 of 76

Node ids are slugify(frontmatter name) per RFC-007:142, but authors link by filename. RFC-007:151 separately requires composes-with markdown links to resolve "by file basename", which needs an index :142 disclaims. The RFC contradicted itself, functionally rather than cosmetically.

Resolution now tries the name-derived id first, then a filename-basename alias index; ambiguous basenames are dropped rather than guessed. Node ids are unchanged and :142 is untouched. Documented in a new Target resolution subsection.

Two RFC contradictions settled

• Dangling bucket is wiki-link everywhere. The prose said wiki-link-dangling, a spelling appearing nowhere else in the 737-line file and disagreeing with both the contract and the kind enum.
• The wiki-link regex was stated twice, non-equivalently. The prose form is canonical. Verified programmatically: the contract JSON-unescaped, the prose span, and the literal compiled into em-graph.mjs are now character-identical.

Verification

Claim	Command	Observed
New suite	node tests/test-em-graph-links.mjs	31 passed, 0 failed
Phase 1 regression	node tests/test-em-graph.mjs	10 passed, 0 failed
Phase 2 regression	node tests/test-em-graph-nodes.mjs	28 passed, 0 failed
Suite genuinely CI-wired	node tests/test-ci-suite-registration.mjs	16 passed, 0 failed
Negative control	node tests/test-em-graph-links.mjs --break-wikilink	exit 1, 11 failures
REQ-8 default output	worktree at 8e6e699, cmp on --hubs and --orphans	byte-identical
RFC anchors	sed -n '239p;251p;314p;320p;332p'	all point where claimed

REQ-8 was checked against the baseline commit directly rather than via the suite's own before/after comparison, which is the proxy layer one short of the claim.

Live corpus: 81 rule nodes carry wiki-link edges (9 dangling), 31 carry composes-with edges (1 dangling). --from pi-minimax-m3-builder-seat --depth 2 returns 5 nodes and 4 edges where 8e6e699 returned count: 1.

Review

Two rounds, both HOLD, all 12 findings dispositioned.

Round 1 on the plan caught a blocker: a step referenced a symbol a later step defined, so verbatim application would have thrown ReferenceError. Round 2 on the frozen diff caught a MAJOR: CRLF line endings silently dropped every composes-with edge — zero edges, zero dangling, exit 0. Latent on this corpus but the same silent-zero signature REQ-6 traps for, and inconsistent with lib/rule-nodes.mjs:67 which already handles CRLF. Fixed with a regression test proven to fail without the fix.

Round 2 also caught a REQ-12 miss: the RFC's Scope "Unbuilt" list still enumerated Phase 2 and Phase 3.

Pre-existing stalenesses fixed in passing

• The mermaid diagram labelled Phase 2 UNBUILT while the phases table and ledger both recorded it SHIPPED.
• slugify._shipped_use claimed the function was unused by the shipped script, though scanRuleNodes has called it since Phase 2.

grep -c UNBUILT on the RFC: 17 before, 10 after.

Deliberately not included

No Implementation ledger row. Its provenance must come from git log --follow on the merge commit, which does not exist yet. Writing a branch-tip SHA there is the defect #582 was filed for and #605 had to repair. It will be filled post-merge.

🤖 Generated with Claude Code

[lantiscooperdev]

Tiered multi-agent review — RFC-007 Phase 3, two rounds, ACCEPT

Covers the full #617 diff (5 files, +1638/-34) across a scout stage, a plan review and a frozen-diff review. Both review rounds returned HOLD; all 12 findings are dispositioned and the actionable ones applied.

Verdict trail

Round	Scope	Seat	Verdict	Blockers
Scouts	RFC anchors, em-graph.mjs runtime state, test/CI wiring, governance + disposition sweep	4 parallel Sonnet	n/a	n/a
1	The plan, with Appendix A frozen	pi / neuralwatt glm-5.2	HOLD → folded	1 BLOCKER, 2 MAJOR, 4 MINOR
2	Frozen diff main...HEAD	pi / neuralwatt glm-5.2	HOLD → fixed in f76dc6b	1 MAJOR, 4 MINOR

Seat spend for the whole arc: $2.71. No Fable seats, no openrouter.

Round 1 findings, closed before any code was written

#	Class	Resolution
1	BLOCKER	Step 1.8's payload called danglingOut(), defined by step 1.9. Verbatim application would have thrown ReferenceError and dropped Phase 2 to 27/28 with no clean in-step fix. Stray call removed.
2	MAJOR	Step 1.4's verify asserted a dangling key nothing emits until 1.9. Rewritten to a loop-presence grep.
3	MAJOR	Step 1.9 expected grep -c 'danglingOut()' → 4 counting the definition, but the definition reads danglingOut = () so the substring never matches it. Corrected to 3.
4	MINOR	Step 3.5's relative UNBUILT count was ambiguous. Restated absolutely: 17 at 8e6e699, 15 after 3.2, 10 after 3.5.
5	MINOR	composesTargets slugified whole prose bullets into garbage refs. Now yields no target.
6	MINOR	Two backticked spans on one bullet: first wins, made explicit.
7	MINOR	Presence-only greps, assessed as back-stopped by the functional suites. No change.

Round 2 findings, closed in f76dc6b

#	Class	Resolution
1	MAJOR	CRLF line endings silently dropped every composes-with edge — zero edges, zero dangling, exit 0. The heading matched (that test trims) but the bullet kept a trailing \r and the bullet regex did not match. Same silent-zero signature REQ-6 traps for, and inconsistent with lib/rule-nodes.mjs:67 which already handles CRLF. Fixed by splitting on /\r?\n/.
2	MINOR	A markdown-link target with a #fragment slugified the fragment into the target. Stripped before basename.
3	MINOR	Section start trimmed, terminator did not, so an indented ## Composes with could start a section an indented ## Next could not end. Both now test the trimmed line.
4	MINOR	RFC Scope's "Unbuilt" list still enumerated Phase 2 and Phase 3, contradicting three other places in the same file. A REQ-12 miss: the first sweep only covered the sites the plan enumerated.
5	MINOR	The two new contract entries lacked the shipped_at field all five older entries carry.

RFC contract changes

• Edge-status rows :84-85 and both contract status fields → SHIPPED, with verified anchors.
• Dangling bucket unified to wiki-link; the wiki-link-dangling spelling appeared nowhere else in the file.
• Machine-readable regex rewritten to match the prose form. Verified programmatically: the contract JSON-unescaped, the prose span, and the literal compiled into em-graph.mjs are character-identical.
• New Target resolution subsection defining the filename-alias index, reconciling the "resolved by file basename" clause with the frontmatter-name node-id rule. :142 itself untouched.

Orchestrator re-verification (independently executed, not taken from seat reports)

• Suites: links 31/31, phase1 10/10, phase2 28/28, ci-suite-registration 16/16. PASS.
• Negative control --break-wikilink → exit 1, 11 failures, and the passing tests are exactly the ones not dependent on wiki-link extraction. PASS.
• REQ-8 against the real baseline: worktree at 8e6e699, cmp on --hubs and --orphans output → byte-identical. The suite's own before/after check is the proxy layer one short of this. PASS.
• RFC anchors 239, 251, 314, 320, 332 each verified to point where claimed, re-checked after the S4 line shifts. PASS.
• CRLF guard proven non-vacuous: the bullet regex returns false against a trailing \r, so the new test fails without the fix. PASS.
• Payload fidelity: every §A.5 block diffed against the plan, verbatim including comments. PASS.

Live corpus behavior

Edge	Nodes with edges	Dangling
wiki-link	81 of 127 rule nodes	9
composes-with	31	1

--from pi-minimax-m3-builder-seat --depth 2 returns 5 nodes and 4 edges; 8e6e699 returned count: 1.

Follow-ups filed (Rule 18 step 9)

• #618 — fill the Implementation ledger row post-merge from git log --follow. Deliberately omitted here: the provenance SHA does not exist until merge, and a branch-tip SHA is the defect #582 was filed for and #605 had to repair. Carries all 5 DEFER fields.
• #619 — em-console.mjs:114 flag whitelist omits --nodes/--edges, so rule/rfc nodes and Phase 3 edges are unreachable through the console. Pre-existing since Phase 2, recorded as an explicit non-goal.
• #620 — an unclosed code fence does not suppress extraction, affecting shipped cites as well as new wiki-link. Design question surfaced by the build, deliberately not decided inside this slice.

CI

8/8 green: validate ×3, substrate, gates-classifiers, p12-invariant-gate, recall-activation, clerk-consolidation.

What user approval is checking

Per Rule 17, your approval in the UI is the review-required gate. Substantive review is complete and both rounds are closed. Worth a specific look before approving: the filename-alias index is a genuine spec extension, adopted because the literal spec resolved 13/132 wiki-links and 0/76 composes-with bullets against the live corpus. It does not change node ids. If you disagree with that call, it is the one decision here that is architectural rather than mechanical.

🤖 Posted by lantiscooperdev (bot account) — Claude Opus 5 (1M context).