gist

The engines were not the problem. The consumer changed. Every tool in this field was built for a human typing a query every few minutes and reading the results with their eyes, who will notice if they look stale. What we have is roughly ten coding agents sharing one working tree, issuing many small searches per session, editing between those searches, and paying for every returned line in tokens. Three consequences fall straight out:

1. Freshness stops being a nicety and becomes correctness. An indexed search that hands back a line which no longer exists has not been fast, it has lied, and an agent has no eyes to catch it with.
2. A set is the wrong shape for an answer. A human scanning two hundred hits finds the definition instantly; an agent reads them in order and burns its context on call sites. Hence --rank, which only ever reorders a set that was already verified.
3. Output is a budget. Every line costs money, which makes terseness a feature and the docs-versus-code partition a real axis rather than a convenience.

So the accelerators here are allowed to decline constantly and never allowed to invent. A persisted trigram index and a resident session can save work; neither may conjure a file set or hand back a stale line, and doubt of any kind returns the query to the live walk. --no-index forces that walk, and it is the differential oracle the indexed path is tested against. The tree tells the truth.

Parity with ripgrep is a constraint rather than a resemblance, because that muscle memory is already embedded in how agents search: the argv, the ignore precedence, the exit codes and the bytes on stdout are ripgrep's, measured on every run against a live rg oracle instead of asserted here. Where gist diverges it is an improvement - identical-or-superset results, better in behavior, performance or robustness, never a regression - or it is a bug. There is no third category, and the eight cases are an enumerable bucket in gist --schema rather than a paragraph that can drift from the binary.

Most of the machinery is inherited and cited: Thompson's linear matcher for patterns that cannot be allowed to detonate, a vendored PCRE2 for the ones that are not regular at all, Cox's trigram construction for deciding what never has to be read. The one piece of mathematics that is ours is the crest sieve, a sound necessary condition on forced class runs. It prunes precisely what the whole trigram family concedes - the literal-free class repetitions like [0-9a-f]{12} that a hash or MAC address hunt is made of - and it is sound by construction rather than by testing, since every term in the bound rounds down and any construct the calculus cannot certify contributes nothing.

Each generation of this tool found a different thing not to look at. Haertel's grep refused to touch every byte; Cox's index refused to open the file. This one adds the refusal they left on the table: not looking again at what has not changed.