fix(syntax): implement Janet backtick long strings

[takeokunn]

Independent of the lint-rule PR chain — this touches only packages/core/syntax.

classify_janet had no backtick arm, so a backtick was neither a delimiter nor whitespace and got absorbed into an atom — which meant a long string's contents were parsed as code. Triple-backtick docstrings are the dominant Janet idiom, so this is not an edge case.

Differential over 241 real .janet files

janet, spork, jpm, circlet, janet-sh — all 241 of which Janet's own reader accepts, so every paredit failure was paredit's.

parse failures      9 → 2      (7 fixed, 0 newly broken)
trees changed            42     (all reductions; 9880 phantom atoms removed)
trees identical         190
top-level form counts vs Janet's own reader:  230/232 → 239/239

The two remaining failures are a different, pre-existing defect I deliberately left alone: ' (quote) is missing from classify_janet's prefix table, so it glues onto a following " (spork/cjanet.janet:31, janet/test/suite-peg.janet:405). Identical on the pre-change binary. It's the same class of bug in another reader macro and it silently mis-shapes every 'foo in every Janet file — worth its own PR.

What parse.c specifies, verified against a 1.41.3-dev build

• A single backtick opens one. root (:643) is unconditional — there is no run-length test.
• The close is exactly N, not at least N. stringend returns 0, re-dispatching the revealing character, so surplus backticks open the next datum: ```ab x` ```` is two values.
• The opener is the whole run, so an empty long string is unwritable — Janet errors on ``.
• Newlines yes, escapes no — the in-string branch is push_buf with no backslash case at all.
• @ shares the extent, as a buffer literal.
• Backtick is not a symbol character (symchars[3] = 0x07fffffe leaves bit 0 clear), so it also ends the token before it. Without that arm in is_atom_boundary, (foo`bar`) would glue. This wasn't in the original bug report and turned out load-bearing.

An unterminated long string is refused as UnterminatedString, matching Janet's unexpected end of source. Reading to EOF as one atom is precisely the silent corruption this removes.

Evidence beyond the diff

Differential fuzz, 2500 cases, same seed both binaries: 1385 agree before → 2500 agree after, including all 332 cases Janet rejects. The first generator was thrown away — it mixed in tokens whose paredit/Janet divergence predates this change, so "both reject" was coincidental agreement rather than evidence.

Round-trip idempotence passes 239/239, which proves nothing on its own (a consistently wrong parse is a fixed point), so Janet itself was the oracle for whether the value survives.

Other dialects unaffected

3167 files, 1414 of them containing backticks, zero differences between pre- and post-change binaries:

corpus	files	with backtick	different
nix store .el/.scm/.lisp	2500	1291	0
nix store .clj .asd .rkt .carp .lfe .hy .fnl .ss	627	123	0
repo tests/fixtures	40	—	0

Plus a unit test pinning backtick as ReaderPrefix::Quasiquote for all ten other dialects, so a future edit to has_long_strings can't quietly widen.

A formatter bug this surfaced, not fixed here

stringend (:355) does indent_col = top.column - 1 — a long string's value depends on the column of its opening backticks. edit format preserves the literal's bytes exactly and moves that column, so 30 files still change meaning:

(defn f

line one
indented

[x] x)

Janet reads the source string as "line one\n indented"; after edit format it reads " line one\n indented\n ".

Net this is still an improvement — 199 → 209 files round-trip value-identically, and exactly one file (jpm/cc.janet) regressed from value-preserving to value-changing. Three that look new simply didn't parse before. And the pre-change formatter failed worse in kind on the same input, collapsing the docstring's newlines outright. It's a formatter defect with its own blast radius and belongs in its own change.

Cost

The dialect test sits before the byte test in is_atom_boundary deliberately: self.dialect is loop-invariant across the per-byte calls made for every atom in the document, so for the other nine dialects it folds away rather than costing a comparison per byte. Nothing new allocates. This is a code argument, not a measurement — load average was 73 from parallel agents, so no local benchmark would have meant anything.

Verification

cargo build --workspace, cargo test --workspace, cargo test --test cli (3083 passed), cargo fmt --all --check, cargo clippy --all-targets --all-features -- -D warnings — all exit 0.

No pinned count or golden moved, consistent with the fix being confined to a dialect no fixture exercises with backticks. The repo's only .janet fixture contains no backticks and formats byte-identically under both binaries — part of why this went unnoticed.