A writer that gives back the file it was given - #45

Closes #12.

parseIdf(text, schema, { preserveFormatting: true }) retains feature 005's syntax layer and an index from each statement to the object it produced. writeIdf then walks the two together: every object nothing has changed is reproduced from the characters it was read from, and everything between the objects is copied.

Blocked on idfkit/idfkit-conformance#4. This branch pins conformance-2026.10 and governance-2026.14, and those tags are cut from that repository's main after it merges. CI here is red until they exist — that is "publish before you pin" working, not a fault in this branch.

Why the diff is small for what it does

The gaps between statements are emitted unconditionally and only the statements are decided. Every character is either inside a statement or in a gap, because the layer tiles the text, so the comments, the blank lines that group four hundred surfaces into rooms, the line endings whatever they are, and the presence or absence of a trailing newline are all copied without a decision. Three requirements fall out rather than being enforced: nothing is reordered, no header is added, and the version statement is not moved.

The read makes one pass. scanIdf and lex now expose their scan handlers and the preserving read composes them, so the option costs the layer's own work rather than a second read of the text. With the option off, nothing is built.

The touched record closes three paths that notified nobody

• The rename. It rewrites every referencing field directly, bypassing the accessor on purpose, so none of those objects said they had changed. A writer trusting the listener emitted them from their original text, producing a file that loads and names a construction layer that no longer exists. The mark is made on both branches of the retarget loop; the second is reached only by a reference held inside an extensible repeat, so a mark on the first alone passes every test whose reference is a plain field.
• In-place extensible mutation. get extensible() handed back the object's own array. It now returns an Array subclass whose mutators are heard, holding repeats whose fields are own accessors. No Proxy.
• Creation, cloning and attachment, which leave an object with no index.

One spelling is still not heard — replacing a whole repeat by index — and it is documented and tested as the limit it is, with splice asserted beside it as the tracked alternative.

Measured, and one number moved because of it

bench/preserve.mjs reports three figures. One of them changed the design: reading a coordinate through an armed repeat cost 33x a plain read over the reference model's 33,012 vertices, which is exactly the price rejecting a Proxy was meant to avoid, reached by another road. Repeats are now armed only when the document carries a retained source — an editing document pays it, a geometry pipeline does not, and the non-preserving path is back to 1.75x.

It also reports a number the plan did not expect: a preserving read costs 1.54x a plain read, not the 1.25x SC-005 states. The two are different claims. The layer's own budget, that scanIdf costs no more than a quarter over a plain read, holds at 0.53x and is gated in bench/budget.mjs. A preserving read does both jobs, so its floor is the sum.

SC-012 was amended

The writer is 47.7 KB of dist and put the install 14,308 bytes over 1.75 MB. The budget moved to 1.875 MiB, in its own commit, with the reasoning in the gate's header: the capability is core, 1.8 MiB would have cleared the measurement by less than the feature cost, and this leaves 114 KB. The lever nobody pulled is recorded there too — 396 KB of the install is source maps, which serve debugging and nothing at runtime.

packages/idfkit/language.js claimed "98.3 KB free". That was three features stale; the real figure when this work started was 32.7 KB, and the plan was written against the wrong one.