perf(router): remove __debug_manifest endpoint + dead trie ancestry (−21% manifest) - #728

What

Removes the ?__debug_manifest query endpoint (an early opt-in JSON dump of the manifest/trie for inspection) and, with it, the dead per-leaf trie ancestry it was the only consumer of. Net effect: the per-router manifest chunk is 21.6% smaller (5.71MB → 4.47MB at 26k routes) on every runtime, with no behavior change.

Why it's safe (nothing user-facing breaks)

The endpoint's only data dependency was TrieLeaf.a — the ancestry shortCodes on every trie leaf. Route matching never reads it (verified: trie-matching.ts and find-match.ts reference leaf.a zero times; the source already flagged it debug-only). So once the endpoint and extractAncestryFromTrie are gone, a is pure dead weight, and removing it can't change matching, href()/reverse(), or any request behavior.

Deliberately not touched:

• The shortCode system (mount indices used in server/context.ts counter keys and the DSL) — that's load-bearing; only the trie-leaf snapshot of it is removed.
• The separate debugManifest() router method (router/debug-manifest.ts) — a different surface that re-walks handlers into a SerializedManifest; left fully intact.

Changes

• rsc/handler.ts — delete the __debug_manifest endpoint block.
• router.ts / router-options.ts / router-interfaces.ts — remove the internal allowDebugManifest option; cache/cache-key-utils.ts — drop it from the reserved-param set.
• build/route-trie.ts — remove TrieLeaf.a, the routeAncestry param from buildRouteTrie, and extractAncestryFromTrie; buildPerRouterTrie's "no ancestry ⇒ no routes" guard becomes an explicit empty-routeManifest check.
• build/generate-manifest.ts — remove captureAncestry, the _routeAncestry field, and the routeAncestry threading through the recursive builder.
• vite/discovery/discover-routers.ts — drop the _routeAncestry merge and the now-redundant guard.
• Tests — the trie test fixtures that carried synthetic a: leaves / passed the ancestry arg are updated (5 mechanical, parallelized); per-router-manifest.test.ts swaps its extractAncestryFromTrie-based route-name assertions for a local trieRouteNames() walk; dev-prod-trie-parity.test.ts now asserts the stronger invariant — the per-router trie is byte-identical across mount indices (nothing embeds the mount index anymore).
• Docs — feature-map.md (drop allowDebugManifest), matching-and-lazy-discovery.md (remove the leaf.a caveat + the "stop serializing leaf.a" backlog item this completes).

Effect

• Manifest payload −21.6% at 26k routes, everywhere (node/vercel/cloudflare) — shrinks the JSON.parse floor, not just compile. Negligible for normal apps (the win scales with route count), meaningful at the tail.
• Per-router trie is now fully mount-index-independent — a strictly stronger dev/prod parity guarantee.

Test plan

• pnpm run typecheck
• pnpm run test:unit:all — 1195+ tests across all packages, incl. find-match, trie-matching, trie-regex-parity, per-router-manifest, dev-prod-trie-parity.
• semantic-matrix 67/67 (dev + production) — middleware scope/ordering/context visibility/PE-JS parity/PPR all intact.
• stress-demo build: manifest 5.71MB → 4.47MB, M0L0R/"a":[ markers gone, real routes 200, genuine miss 404.
• pnpm run lint, pnpm run format, pnpm check:e2e-bucketing.

Notes

• Independent of perf(vite): ship route manifest as a workerd Text module on cloudflare builds (refs #665) #710 (the Cloudflare Text-module manifest); they compose — this shrinks the bytes, perf(vite): ship route manifest as a workerd Text module on cloudflare builds (refs #665) #710 changes how they ship.
• allowDebugManifest was @internal (not documented public API), so removal follows the pre-release "remove, don't deprecate" hygiene rule.