Beagle SCM

DOG-050: LINK.rl fuses search and symbol anchors — :'text', :"snippet", :/re/, :name

BEE-073 rules four new anchors after a link target's :. LINK.rl knows digits only, so CAPO.c:'Compose URI' fuses as CAPO.c and the anchor falls into prose. Extend anchor, make commit disjoint from line, give long anchors a cap of their own, and ship a ragel splitter so no consumer peels a ref by hand. Seven includers regenerate their twins.

Input

Context

• tok/LINK.rl:88:GG line, tok/LINK.rl:89:GG commit, tok/LINK.rl:90:GG anchor, which relanch and ticket take.
• Digits are ron64, so :30:4711 matches line and commit both; scanner order decides today.
• Includers: FREE, MDT, MKDT, MKDTB, RSTT, TOKL, TXTT (tok/*.c.rl); each binds the parts and owns a generated twin.
• The 10-char anchor cap and the TOKUNCUT retreat: tok/FREE.c.rl:118:yQ, tok/MDT.c.rl:108:6z, tok/TOKL.c.rl:53:IB.
• Consumers split the fused F themselves today (bee door.js:517, by lastIndexOf(":")); a body with a colon breaks that.
• seg never starts with a dot and rooted heads are / or ./ only (tok/LINK.rl:94:GG): no .c scope, no //repo head.

Goals

• grep = ":" "'" body "'", spot = ":" '"' body '"', regex = ":" "/" body "/"; body = ([^delim\\\n] | "\\" any)+.
• sym = ":" [A-Za-z_] [A-Za-z0-9_]*, disjoint from line by its first byte.
• commit = ":" [0-9]{1,10} ":" (ron64{2,10} - [0-9]+): a hashlet carries a non-digit, so :30:4711 is line:col only.
• The delimited anchors and line/commit vouch for relanch and rooted; sym rides relknown and ticket only.
• Caps: a delimited body up to 64 bytes, a symbol up to 32; over the cap the ref retreats exactly as today.
• An unclosed delimiter fuses no anchor: main.c:'open stays main.c plus prose.
• query = grep | spot | regex | sym; scope = dirs? ext query | dirs query: .c:/re/, abc/.c:'x', abc/:'x' (BEE-073).
• A scope's ext faces the KNOWN gate as relknown does; a scope never takes line or commit.
• authed = "//" seg "/" (relanch | scope): one seg of authority, then a path, then the mandatory anchor.
• The splitter reports authority and scope (dir, ext) beside the rest, so a consumer tells a query from a reference.
• A splitter machine over one fused ref: target, kind (line, perma, grep, spot, regex, sym), line, col, hashlet, body span.
• Repro first (work 3): a fusing case per new anchor and per includer family, then the non-fusing ones.
• Non-fusing: either/or:pick, 12:30, example.com:http, main.c:'open, a body past the cap; x.c:30:4711 fuses as line:col.
• Non-fusing heads: //localhost:8080, https://foo.io/bar.c, http://host/a.c:12, //repo/x.c, a bare .c, .c:30.

Constraints

• Ragel only, no byte scanning in C (abc); the existing LINK tests stay green and unedited.
• The F token covers target plus anchor whole (DOG-043); the header comment of LINK.rl stays the spec of record.
• tok/INDEX.md gains the splitter entry (index); comment budgets per comment.

WIP

Design decisions

• Machine difference (ron64{2,10} - [0-9]+) over scanner order: the two anchors stop overlapping in the grammar itself.
• sym does not vouch: a/b:word is common prose, while a quoted or slashed body after a colon is not.
• No stars in a scope (BEE-073): the ext family already means any depth, and */ would close the comment a link sits in.
• A scheme:// run must stay prose: no includer has a URL rule to eat it first, so the guard is part of this ticket.
• authed takes authsym too (review, 2026-09-18): without it //libdog/HUNK.c:sym lexed HEADLESS, resolving elsewhere.
• One on_gated action serves a scope and an authsym: both rest on the KNOWN gate and both must survive the cap.

TODOs

• repro tests, fusing and non-fusing
• LINK.rl: the four anchors, the disjoint hashlet, the vouching split
• the caps in the seven includers, twins regenerated
• scope targets and the // head, with the scheme:// guard
• the splitter machine and its INDEX.md entry

Blockers and bummers

• file.c:k4:d8K3 and DOG-032:k4:d8K3 read :k4 as a SYM anchor now, so the dead ron64 offset fuses that far.
• gritzko ruled those four expectations amended FPSPS -> FPS; the one pre-existing test this ticket touched.
• The review caught a spec gap: authed had no authsym, so //libdog/HUNK.c:sym lexed headless and resolved elsewhere.
• The scheme:// guard eats the slashes only: http://host/a.c:12 still fuses its inner host/a.c:12, as it does today.
• Eating the whole URL was refused — {"https://foo.io/bar.c", "SPPPSSSPF"} pins bar.c fusing, and no test may be edited.
• Likewise //repo/x.c fuses repo/x.c on its known ext; only the AUTHORITY reading is blocked, which is what was asked.
• TOKUNCUT still trims when a ron64 byte follows a CLOSED delimiter, so x.c:/re/g retreats; harmless, but not exact.
• tok/INDEX.md does not exist; the splitter's entry went to the tok section of dog/INDEX.md instead.
• No fuzz target is configured in dog/build (gcc, no libFuzzer), so the short fuzz run was skipped.
• The baseline was 85 tests, not the 82 the brief expected; 85 passed before the change and 85 after.

Outcome

Working tree only, nothing committed: tok/LINK.rl grew the anchors, the scope, authsym and the authority head; tok/LINK.h + tok/LINK.c.rl are the new splitter; the seven includers and tok/TOK.c bind them; twins regenerated, and all 85 tests pass.