Harden the compression hook and refresh the published numbers

[lightningpixel]

Pre-release pass. Everything here came from running orvix against a real project rather than against its fixtures.

The hook could fail a turn

Every guard inside orvix-compress.mjs answers an unusable input by leaving the original result standing. An exception escaping main() did the opposite: non-zero exit, error in the agent's transcript.

It is reachable in practice — a .orvix/symbols.json written by another version of orvix parses as JSON but need not hold the shape this version expects:

TypeError: ids.map is not a function
    at compressRead (hooks/orvix-compress.mjs:187:6)
    at main (hooks/orvix-compress.mjs:56:23)

The hook runs on every tool call. It now fails the way the rest of the file already does.

orvix init was silent about the saved choice

Applying an answer given in another repository months ago, with no indication of where it came from, reads as the tool guessing:

using saved choice: claude (C:\Users\MaXx_\.orvix\config.json — override with --agent)
created CLAUDE.md
…

The published numbers had drifted

The README opens on "Every number below comes from a command in this repository. Re-run them." That makes stale figures the one defect a reader finds without trying.

	was	now
index (README:29)	389 symbols, 1 127 edges	395 symbols, 1 150 edges
measure	~1,912 / ~33,254 / ~1,990	~2,000 / ~33,630 / ~1,998
bench	37.4 / 17.2 / 54.6 ms	34.7 / 16.7 / 51.4 ms
scale	274 / 790 / 2 825 ms	285 / 746 / 2 547 ms
tests (README:91)	362	370
tests (AGENTS.md:11)	246	370

All four commands re-run, not copied. benchmark and the 31% hot-cache share were unchanged — benchmarks/results/synthetic.json differs only by its timestamp, so compression behaviour is byte-identical to before the recent fixes.

Newly documented

Calls dispatched through a string are invisible to the call graph — an Electron IPC channel, an event name, a route table, a DI container. Found by watching an agent ask who-calls getSettings, get the 9 main-process callers, and then correctly fall back to grep for the renderer half behind ipcMain.handle('settings:get'). This gap is larger than the ones the section already listed.

AGENTS.md also now records the language-family constraint as an invariant, so it is not removed by someone who does not know why it is there.

Verification

• npm run build, npx tsc --noEmit, npm test — 370 tests pass
• npm run benchmark — ratio 0.699, success_delta 0, both fixtures preserved
• Both new tests written first and watched fail: the hook test with the real TypeError, the init test with output that mentioned neither the agent nor the config file
• orvix init run for real against the global config, not only under test

Not in this PR

The PostToolUse:Grep hook warning is still undiagnosed. Every successful Grep compression produces one; the Bash compression produces none. Until that is settled, the Grep share of orvix gain is unverified — 88% of the total on the session where it was observed.