libexpr: builtins.tecnixQueryFile, and fingerprints that record observations

[joshheinrichs-shopify]

Replace the three World-specific manifest builtins from #72 (tectonixManifestEntry, tectonixManifestKeys, tectonixManifestIdToPath) and their .meta/manifest.json# synthetic keys with one generic builtin that queries any JSON or TOML file in the repository, and with a fingerprint format that records what was asked and what the answer was. Tecnix no longer knows anything about World.

builtins.tecnixQueryFile {
  path = ../../.meta/manifest.json;   # format inferred from the extension
  op = "get";                         # get | has | keys | find
  at = [ "//areas/tools/dev" ];       # object keys, root down
  # find only: where = "id"; value = "W-123456";
}

The model. A closure maps a repo path to a fingerprint, and a fingerprint records what evaluation observed about the path: git:<oid>;mode=… is the object observed there, absent that none was. A target that only queried a data file observed less than the whole object, so it records the answers it got:

".meta/manifest.json": "query:json\tget\t//areas/tools/dev\t{\"id\":\"W-c10e84\"}\njson\thas\t//areas\ttrue"

one line per distinct query, sorted. The key stays the repo path, so everything that only knows paths keeps working (tec's affected-targets stays conservatively correct with no change), and the TXDC row format is untouched. A target that also read the file whole records the oid, which subsumes the queries. #72 had moved the observation into the key instead, which is why it needed a key grammar, a # special case in validation, and a tec change.

How it works. The builtin reads the file unrecorded through the repo accessor, parses it once per evaluation, evaluates the op, and records an access <file>\0<line> through the ordinary recordAccess, so labels, tecnixMemoize replay and thread confinement apply unchanged; the line is built in thread-local scratch and interned on first sight like a path. Flattening groups accesses by file into one query: fingerprint. Validation (dependencyFingerprintHolds, replacing the string equality in candidateMatches) splits a stored query: fingerprint into views, re-asks each line of the file as the current tree serves it (parsed once per thread per run, read unrecorded), and memoizes per (path, line) in the existing thread-local memo. Evaluation and validation run the same evaluateDataQuery on the same parsed tree, so they cannot disagree. A line that does not parse, an unknown op or format, or a file that no longer parses does not hold: a miss, never a stale hit. A missing file answers every query as an empty document would, and that answer is recorded.

Format: a line is <format>\t<op>\t<fixed args…>\t<at segment…>\t<answer>, the answer last so a variable number of segments is unambiguous. Fields never contain tab, newline or NUL: the builtin rejects string arguments that do, and answers are absent, compact JSON when at most 64 bytes (which escapes control characters), or sha256:<hex> of it. Parsing a fingerprint is splitting, with no unescaping.

Outside the repo accessor the builtin is fromJSON (readFile path) plus the selection; a path pure evaluation forbids throws as readFile would (lstat, not maybeLstat, which reports forbidden paths as absent).

Removed: the three builtins, getTecnixManifestJson and its EvalData fields, manifestPathFingerprint and the per-thread parsed manifest in eval-cache.cc. The unsafeTectonixInternalManifest* builtins stay; they record the whole file, correctly. Old cache rows (keys .meta/manifest.json#…) read as paths that do not exist and miss; no blob version bump is needed.

Tests: src/libexpr-tests/tecnix-data-query.cc (27: op semantics, line round-trips, rejected characters, answer hashing, TOML and JSON of one tree agreeing, every hold/miss case including unparseable and unknown lines, grouping and subsumption in dependencyFingerprints); tests/functional/tecnix/data-query.sh (replaces manifest-tracking.sh, keeping its memoize cases): exact fingerprints per target; across ten commits, adding a zone misses only keys and the whole-file reader, has b survives b's renumbering, a comment-only TOML edit misses only the whole-file reader, a file whose absence was observed misses when it appears, a dirty data file is queried from disk and its observation follows; argument validation; the builtin outside tracking; the older-rev read. The tecnix suite and the Tecnix unit tests pass.

Docs: explainer §4.2 and the fingerprint vocabulary; a guardrail on fingerprints as observations.

Also, a second commit: EvalSettings::getDefaultNixPath no longer throws when a channels directory cannot be stat'ed. It used the throwing std::filesystem::exists, so the EPERM a strict Darwin sandbox gives for /nix/var/nix/profiles aborted every impure EvalState: every libexpr test fixture, the C API flake tests and the inner nix of recursive.sh, which is how this pin failed to build its own test suites on a Mac with sandbox = true (macOS defaults it off). World's Darwin build of this pin passes all suites with it (unit tests and 200/200 functional).