[diegosouzapw]

Refs #13455

Root cause

Issue #13455 bundled three distinct symptoms with three different fixability profiles:

1. <system-reminder>/envelope destruction — already fixed by fix(compression): preserve instruction blocks from lossy engines #13523 (landed one day
   before this analysis), unrelated to this PR.
2. Boundary whitespace lost around preserved inline segments (backticks, URLs, etc.)
   in plain prose — this PR's fix. compressMessageText() in
   open-sse/services/compression/engines/llmlingua/index.ts splits a message into
   alternating prose/preserved segments and re-joins them with a bare parts.join("")
   after each prose segment is sent through the compression backend. The real ONNX
   worker backend rebuilds output from surviving wordpieces after tokenization, which
   trims/normalizes segment boundaries as a side effect — it never promised to preserve
   the caller's leading/trailing whitespace. Whenever a prose segment sat directly
   against a preserved placeholder, the bare join glued the two together with no space
   ("NEVER run \rm -rf` on the target host" → `` "runrm -rf`target host" ``).
3. Case flattening / negation-word dropping inside surviving prose — this is the
   default TinyBERT checkpoint's own per-token pruning decision, made inside the
   third-party @atjsh/llmlingua-2 ONNX model with no token-span/offset-mapping API to
   hook into. Not fixable purely in OmniRoute's own code without either pre-tombstoning
   individual negation words (fragmenting prose for every stacked engine, unreviewed
   blast radius), switching the default model to an unverified cased checkpoint, or
   forking the third-party reconstruction logic — each a real design decision the owner
   needs to make. Left for a follow-up rather than bundled into this fix.

Fix

compressProseText() now captures the original segment's leading/trailing whitespace
before calling the backend, and reinserts that exact original whitespace run (not a
synthesized single space) around the backend's result if the result doesn't already carry
equivalent boundary whitespace — so a segment that bordered a newline stays on its own
line instead of collapsing onto one. The accept/reject "did this actually get shorter"
gate in the same function now compares using the post-restoration length, and rejects a
segment whose only change would be reattaching its own stripped whitespace (no real
compression happened).

Scoped entirely to open-sse/services/compression/engines/llmlingua/index.ts — no
changes to preservation.ts (the already-shipped #13523 envelope fix) or any other
engine.

Regression test

tests/unit/compression/llmlingua-whitespace-restitch.test.ts (5 cases: the already-fixed
envelope regression guard, the reported glued-segment shape, a preserved block at the
very start of the message, one at the very end, and a newline boundary before a fenced
block).

RED on unfixed code (4/5 failing):

✖ preserves boundary whitespace and does not glue prose to a preserved inline-code span
✖ keeps a preserved segment at the very start of the message glued to nothing extra (no leading prose)
✖ keeps a preserved segment at the very end of the message glued to nothing extra (no trailing prose)
✖ preserves a newline boundary before a fenced block instead of collapsing it to a space
ℹ tests 5 / pass 1 / fail 4

GREEN after the fix (5/5):

✔ [already fixed by #13523] content inside a <system-reminder> envelope survives byte-identical
✔ preserves boundary whitespace and does not glue prose to a preserved inline-code span
✔ keeps a preserved segment at the very start of the message glued to nothing extra (no leading prose)
✔ keeps a preserved segment at the very end of the message glued to nothing extra (no trailing prose)
✔ preserves a newline boundary before a fenced block instead of collapsing it to a space
ℹ tests 5 / pass 5 / fail 0

Existing llmlingua suites re-run clean (no regression): llmlingua-failopen.test.ts,
llmlingua-ultra-entry.test.ts, llmlingua-model-store.test.ts,
llmlingua-worker-resolution.test.ts, llmlingua-worker-spawn-12822.test.ts,
llmlingua-worker.test.ts, ultra-slm-tier.test.ts — 29/29 pass, 0 fail.

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — clean (0 errors).
• npx prettier --check <changed files> — all matched files use Prettier code style.
• npm run check:open-sse-typecheck — OK, 0 errors, all within frozen baseline.
• node scripts/check/check-file-size.mjs — no new ✗ on touched files; the pre-existing ✗ entries (open-sse/utils/stream.ts, open-sse/handlers/imageGeneration.ts, tests/unit/batch_api.test.ts, tests/unit/image-generation-handler.test.ts) are base drift, untouched by this PR.
• node scripts/check/check-complexity.mjs — OK, 2826 violations vs baseline 3218 (ratchet respected).
• node scripts/check/check-cognitive-complexity.mjs — OK, 1276 violations vs baseline 1437 (ratchet respected).
• node scripts/check/check-test-discovery.mjs — OK, new test file discovered (5799 test files, 32 collectors).

Not covered here

Per the plan-file's explicit scope decision, this PR fixes ONLY the boundary-whitespace
loss described above. It does not address:

• Case flattening (e.g. SSH → ssh) or negation/stopword dropping within surviving
  prose text — that is the default TinyBERT checkpoint's own token-pruning decision made
  inside the third-party ONNX model, with no per-word "never prune this" hook exposed
  the way the rule-based ultra engine got in fix(backend): ultra heuristic prunes polarity/negation words and collapses newlines, inverting instructions #13454. Fixing it needs either mapping
  output back onto the original text's exact character spans, or switching the default
  to a verified cased checkpoint — both real design decisions outside a single-issue
  bug-fix PR. Recommend tracking as its own follow-up issue.

[diegosouzapw]

Closing in favour of #13768 by @lorenzozanee, which covers the same llmlingua fidelity area for #13455 and was opened first.

This PR came out of a parallel maintainer sweep that was already running when #13768 appeared. The contributor's PR keeps the credit and goes through review.

One thing worth carrying over into the review of #13768: the core symptom in #13455 is boundary whitespace disappearing around preserved spans (run \rm -rf` on → ``runrm -rfon``). The 5 model-free regression cases written here — whitespace around inline code mid-line, a preserved span at the very start and at the very end of a message, a newline before a fenced code block, plus the ` envelope guard from #13523 — are a good check to run against that branch.