[Kinflou]

Step 2 of 3 of the compatibility work (#119, ComlineProject/runtime#15): what each frozen version records and provides for comline@2.

Package id

• id = "<32 lowercase hex>" in config.idp (128 random bits; std's randomly seeded hasher, the clock and the pid through blake3, no new dependency). Checked when config.idp is interpreted.
• package::id::{generate, check, add_to_config_source, ensure_in_config}: the CLI's new writes one, build adds one to a config.idp without it, inserting a single line after congregation.
• Changing the id is PackageIdChanged, major (identity). Adding one where there was none is no change.

comline_core::compat (behind compiler)

• facts(project, version) -> PackageFacts { id, version, protocols: [{name, line, shape_hash, functions: [{name, since}]}], encodings: [{kind, name, params, param_hash}] }, serde-serializable; also facts_of_commit and facts_of_context.
• Keyed by qualified path (namespace::Protocol): protocols of the same name in two schemas have their own facts. ProtocolFacts carries path, namespace and name. Each protocol's changes come from the diff with every other protocol left out, and its reach is taken from its own schema. A renamed namespace or protocol keeps its history (line and sinces).
• Line: the release at which a protocol last broke. Computed against the last release with step 1's compare: a breaking change counts for a protocol when it concerns the protocol or a declaration it reaches (diff::protocol_reach, diff::concerns_protocol; dependencies by reach; thrown errors and constants named by validator arguments included). Namespace and protocol renames don't count, since names don't reach the wire. A change to the package's wire.format, wire.framing or handshake moves every line.
• Since: the release a function first appeared in, per protocol. If a function is removed in one release and added back in a later one, its since starts again.
• Releases only: before 1.0 a break is the 0.x minor. A build carries the line and since values of the release it leads to (0.2.0-dev.2 → 0.2.0). If comline release makes a different version (a raised bump, or --stable), it renames them in a tree of its own. The values are computed against the last release, not build by build.
• Storage: a facts blob in each build's root tree, beside config and docs. Schema blobs and schema_ir_hash don't change. The no-change check ignores the blob, so an old store doesn't produce an empty build. A commit without the blob gets its facts computed on read. A test recomputes every commit's facts and compares them with the stored ones.
• Shape hash (computed on demand): blake3 of canonical CBOR over the encodings in play (by parameter hash), the functions (name, arguments, return, thrown errors, @provider/@consumer) and every declaration they reach. Aliases are resolved, and declarations are numbered in the order first reached, so no type name or namespace goes in. Order counts only where it travels. Function order counts where the framing calls by index (calls), and field and variant order where the format goes by position (struct_fields, enum_variants); otherwise they're sorted by name. An encoding without an official profile counts as order-sensitive. Tests: a function reorder keeps the hash under jsonrpc and changes it under datagram@2, and a field reorder keeps it under json@1 and changes it under packed@1. Docstrings, spans, validators, constants, defaults and advisory annotations are left out. Tests show that docstring, span, namespace, alias and constant edits leave the hash alone, while field, argument, function-name and thrown-error edits change it, and only for the protocol that reaches them.
• Encoding parameter hash: blake3 of CBOR (kind, name, params), covering each format, framing and handshake in play. A declared encoding's profile hash counts as its profile parameter.

For step 3: identifying a protocol on the wire

Facts are keyed by the qualified path, but the shape hash leaves the namespace out, as decided. How comline@2 names a protocol in its offer is step 3's question (runtime#15). It could send the path, which would break peers on a namespace rename even though the namespace doesn't travel today. Or it could send the bare name plus the shape hash, which is ambiguous when two schemas have a protocol of the same name.

core#156

A field's effective validators are compared with each call's arguments resolved against the validator's parameter defaults. A narrowed default now reports validation_tightened on every use that relies on it, and writing a default out is no change.

Not here: core#155

@open doesn't exist in the language yet: the grammar's enum takes no annotations, and runtime#17 is still open. Freezing enum annotations needs a grammar change plus an IR field, so it stays open.

Refs #119, ComlineProject/runtime#15. Closes #156.