Probe: file-backed social graph (relational data as values) — 2026-10-18

Model: Muse Spark 1.3 Contributor (same family as the maintainer's assistant; the external-model route was unavailable — the opencode API key is expired and openai-codex is rate-limited — so the probe ran as a fresh agent with no compiler, internals, RFCs, or prior findings in context, in a stripped tree). Host: Linux x86_64. Compiler: tychoc 0.8.7.

Program: graph.ty (432 lines, single file, no imports; thrown away after this record). A social network as one owned Graph value ([int: Person] + next_id): add/rename/remove people, friendships, friends-of-friends and BFS shortest-path queries, save/load with a byte-identity round-trip check, clean diagnoses for a missing file and three corrupt-file shapes. MUST-use coverage: int-keyed maps, arrays, Option/Result with both or_return and both or_else spellings, a generic find_pos over [$T], a custom subscript person (read + in-place rename).

What went right

Most of the documented language worked on first use: struct value semantics, inout through a map place (push(g.people[a].friends, b)), insertion-ordered keys(), exhaustive match on Option/Result, or_return chains in the loader, generic inference with == and no where clause, f-strings with call holes, Result(void, string) with bare Ok(), the subscript one-use rule, Option/None typing from context. The tutorial's "everything is a value" framing made the index-pool design feel like the intended shape rather than a workaround. First full-program compile needed exactly one fix.

Findings (adjudicated against main, all reproduced)

1. Missing composite map read segfaults (exit 139, silent), both compilers. m[99].name on [int: Person] with no key 99. maps.md:39 promises a deep copy of the value type's zero. Scalar misses correctly return zero. UBSan: non-zero offset applied to a null pointer. Minimized to 6 lines; filed and fixed as FRICTION 204 (NULL guard in tycho_str_copy, shared runtime). The probe's in-guards and present-keys-only subscript rule were load-bearing against this bug.
2. or_else refused in void main, with a good message pointing at or_return or match. Documented (spec/10-statements.md:219). Shaping friction only: unwraps moved into helpers, main a thin match/die wrapper.
3. split("", ",") returns [""]. Follows from the spec's "n separators yield n+1 fields" (spec/16-builtins.md:136) but is easy to misread; the loader special-cases empty fields. Minor.
4. Unused functions are warnings; the brief's zero-warning rule made every helper earn its place. Working as designed.
5. read_file fails open ("" for missing/unreadable/empty alike, builtins.md:99); the fail-closed core:io.read_text needs an import. Documented tradeoff, genuine friction for single-file programs.
6. 0 - 1 cargo-culting from examples/ and corelib/ instead of unary minus. Stale style post-0.8.7 (negative literals now adapt); examples lag the language.

Where the language reorganized the problem

No references ⇒ the graph is an index pool by construction; removal is "delete the entry, sweep every friend list." No pop-front/deque ⇒ BFS uses an array queue with a head index. Copies on return ⇒ queries hand out deep copies, mutation only via inout or the subscript place. Strict parsing is manual labor (no exceptions; lax parse fails open per quiet-results.md) ⇒ the loader validates magic header, count agreement, id positivity/uniqueness, name legality, and edge dangling, each with its own message.

Verification

Rebuilt clean from a scratch directory under ./tychoc and ./tychoc1: byte-identical output (33 lines); ASan+UBSan build identical, no findings. m[99] on a missing scalar key returns 0, exit 0, on both compilers.