Self-hosting

A language that can describe its own compiler has demonstrated the parts that matter: enough expressiveness to hold real programs, enough precision that two implementations agree, and a specification complete enough to write a second one from. That is the goal of phase P3 — "Tripon made with Tripon" — and it is being reached one pass at a time, with no hand-waving between the steps.

The three ports

Each pass of the bootstrap compiler is ported to .tp, and each port is held to a frozen oracle produced by the C implementation.

How the oracle works

tests/selfhost.sh builds the port, then runs it and the bootstrap command over the same corpus — every file in examples/ and tests/cases/, plus the port's own source. The file's exact bytes are handed to the port through $TP_SRC, with a sentinel character appended and stripped so that trailing newlines survive a shell round trip that would otherwise eat them. For the lexer, the comparison is the complete token stream:

ralt tokens examples/hello.tp
TK_USE 4 1
TK_TIER 4 5
TK_IDENT 4 10 core
…
TP_SRC="$(cat examples/hello.tp)" ./tp_lexbin   # the same stream, from Tripon

Byte-identical output, or the test fails and says which file. Nothing is normalised, nothing is sorted, and nothing is ignored: a difference in whitespace between two token lines is a failure. That is the point — an oracle that tolerates small differences proves nothing about the second implementation.

The parser oracle compares the AST dump the same way, over 48 files: the 46-file corpus plus lexer.tp and parser.tp itself, which is what makes a parser port meaningful rather than incidental. The port parses its own source, and the dump has to match what the C parser produces for it.

What the port exposed

Writing the same pass twice is a good test of the first one, and it found two real bugs in the bootstrap code generator:

• An else if chain dropped statements after the second level. The ST_IF emission was not recursive, so anything following a nested else if disappeared from the generated C.
• a || (b && c) regrouped. Parenthesised operands were emitted without the precedence parentheses they needed, so the C compiler read the expression as C associates it rather than as it was written.

Neither was a crash. Both were the kind of quiet difference in output that only shows up when a second implementation has to agree with the first.

What the port forced

Tripon v0 has no module-level mutable state and no array parameters, so the typechecker port threads a by-value Compiler struct through every recursive function — state flows in as an argument and out as a return value. That is a genuine constraint of the language as it stands, and the port is a live record of what it costs: more parameters, more returns, no hidden state. It is also an argument in the language's favour — with the struct threaded explicitly, the checker has no ambient state to reason about, and the same discipline is what makes each pass a function of its input.

The flag day

The end of P3 is a single test: a Tripon compiler, compiled by the bootstrap ralt, compiles itself and passes the whole suite — after which the bootstrap can be retired to a bootstrapping reference. Until that happens, each pass is a measurable fraction of the way there, and the oracles above are the ruler.

Why this is worth the effort

Self-hosting is not a purity exercise here. A second implementation in the language being defined is the strongest available test of three properties at once: that the grammar is complete enough to express a compiler (no missing constructs), that the semantics are precise enough for two implementations to agree (no undefined corners), and that the diagnostics are stable enough to be a contract. All three are the same properties an agent writing Tripon code depends on.