docs: add vecr-compress integration provider page

src/oss/integrations/providers/vecr-compress.mdx

@@ -0,0 +1,77 @@
+# vecr-compress
+
+[vecr-compress](https://github.com/h2cker/vecr) is an open-source LLM context compressor with a deterministic retention contract. It pins structured tokens (order IDs, dates, URLs, emails, code references) via an auditable regex whitelist before running token-budget packing. Filler phrases are hard-dropped; remaining sentences are scored by entropy and structural signals (digits, braces, capitalization) and packed greedily into the token budget.
+
+## Overview
+
+| Property | Value |
+|---|---|
+| Package | `vecr-compress` (install with `[langchain]` extra) |
+| Import path | `vecr_compress.adapters.langchain` |
+| License | Apache 2.0 |
+| Python | 3.10+ |
+| Retention contract | Deterministic (regex whitelist, 13 built-in rules) |
+| Streaming | No (one-shot, synchronous) |
+
+## Installation
+
+```bash
+pip install 'vecr-compress[langchain]'
+```
+
+> The older `langchain-vecr-compress` shim package (0.1.0) re-exports this same adapter and is now deprecated in favour of the `[langchain]` extra on the core package. Existing installs keep working; new projects should use the extra.
+
+## Usage
+
+```python
+from langchain.messages import AIMessage, HumanMessage, SystemMessage
+from vecr_compress.adapters.langchain import VecrContextCompressor
+
+compressor = VecrContextCompressor(budget_tokens=2000)
+
+compressed = compressor.compress_messages([
+    SystemMessage(content="You are a refund analyst."),
+    HumanMessage(content="Order ORD-99172 placed 2026-03-15. Amount $1,499.00."),
+    HumanMessage(content="What is the refund status?"),
+])
+```
+
+`AIMessage` objects with `tool_calls` are preserved verbatim and round-trip intact.
+
+## Advanced usage
+
+Access the full compression report for telemetry:
+
+```python
+result = compressor.compress_with_report(messages)
+print(f"Ratio: {result.ratio:.1%}, pinned facts: {len(result.retained_matches)}")
+```
+
+### Opt-in question-aware scoring (v0.1.3+)
+
+For natural-language QA workloads (long prose contexts, real user questions), enable `use_question_relevance=True` to blend Jaccard question overlap into the default heuristic scorer. On the HotpotQA dev probe (N=100, distractor split) this lifts supporting-fact survival at ratio 0.5 by **+9.9 percentage points** (58.0% → 67.9%). Off by default so the deterministic retention contract stays the loud promise. See [BENCHMARK.md](https://github.com/h2cker/vecr/blob/main/docs/BENCHMARK.md#hotpotqa-spike--where-the-synthetic-bench-hits-its-ceiling) for methodology.
+
+```python
+compressor = VecrContextCompressor(
+    budget_tokens=2000,
+    use_question_relevance=True,
+)
+```
+
+The last `HumanMessage` is auto-picked as the question; override with the `question=` kwarg on `compress_messages()` if needed.
+
+### Custom retention rules
+
+```python
+import re
+from vecr_compress import RetentionRule, DEFAULT_RULES
+
+rules = DEFAULT_RULES.with_extra([
+    RetentionRule(name="ticket", pattern=re.compile(r"\bTICKET-\d{4,8}\b")),
+])
+compressor = VecrContextCompressor(budget_tokens=2000, retention_rules=rules)
+```
+
+## API reference
+
+See the [vecr-compress GitHub repo](https://github.com/h2cker/vecr) for full API docs, the retention contract specification ([RETENTION.md](https://github.com/h2cker/vecr/blob/main/RETENTION.md)), and changelog.