diff --git a/CONTEXT.md b/CONTEXT.md
new file mode 100644
index 0000000..429dbf1
--- /dev/null
+++ b/CONTEXT.md
@@ -0,0 +1,188 @@
+# CONTEXT
+
+Orientation for anyone, human or agent, working in this repository. It is the
+vocabulary and the house rules, not the status. Read it before naming a
+function, a test, or an EXPLAIN line, so that new code says the same words the
+existing code says.
+
+Other documents outrank this one where they overlap, and each owns a different
+question:
+
+| document | owns |
+| --- | --- |
+| `design/NATIVE_FORMAT_AND_INTERFACE_SPEC.md` | what the format and the interface ARE |
+| `design/ROADMAP.md` | what is planned |
+| `CHANGELOG.md` | what changed, per release |
+
+This file owns the words.
+
+There is also a `HANDOFF.md`, the running continuity record: what has happened,
+what is in flight, and what to pick up. **It is deliberately not in the
+repository.** It is excluded per clone through `.git/info/exclude`, so a fresh
+clone will not have one and nothing here should be written to depend on it. If
+you have it, read it; it is the single most useful file on the machine that has
+it.
+
+## What this is
+
+pgColumnar is a columnar storage table access method for PostgreSQL, written as
+a clean-room MIT-licensed implementation. It reads and writes its own native
+format, PGCN v1. The extension, the schema, and the access method are all named
+`pgcolumnar`; a table is created `USING pgcolumnar` after
+`CREATE EXTENSION pgcolumnar`. The clean-room record is `PROVENANCE.md`.
+
+It is an extension, and that is a hard constraint rather than a description.
+**Only core WAL mechanisms and existing record types are available.** A design
+that needs new WAL semantics is rejected, not deferred.
+
+## The storage vocabulary
+
+Section 2 of the spec is authoritative. The short form:
+
+- **Storage id**: 64-bit identifier tying a relation to its columnar storage and
+  its catalog rows. Most metadata is keyed by it.
+- **Row group**: a horizontal partition holding up to a configured number of rows
+  across all columns. A relation is a sequence of row groups. It is the write
+  unit and the parallel-scan unit.
+- **Column chunk**: one column's data within one row group.
+- **Vector**: a fixed run of 1024 values inside a column chunk. The unit of
+  decode, of data skipping, and of vectorized execution.
+- **Page**: the on-disk container of one column chunk's encoded vectors. A
+  contiguous byte range in the relation's main fork, which is why the buffer
+  manager, WAL, and page checksums apply.
+- **Zone map**: the Small Materialized Aggregate for a vector or a column chunk.
+  Minimum, maximum, null count, value count.
+- **Delete vector**: the per-row-group bitmap that makes a delete a metadata
+  write rather than a rewrite. Reads merge it (merge-on-read).
+- **Projection**: a secondary physical ordering of a subset of columns, with its
+  own storage, that the planner may scan instead of the base relation.
+- **Row number**: a 1-based logical position of a row within a relation, stable
+  for the life of the row, mapped to a synthetic item pointer for the executor
+  and for indexes.
+
+## Words that do not line up, and will mislead you
+
+These are measured, not remembered. Each one has cost somebody time.
+
+**A stripe is a row group.** The spec says row group. The GUC, the per-table
+option, and the older code say `stripe`: `pgcolumnar.stripe_row_limit`, default
+150000, is "maximum number of rows per stripe" and it is what sets the row-group
+size. The spec's intended `row_group_limit` (target 122880, a multiple of the
+1024 vector length) is the newer name for the same thing. Both appear in the
+tree. Prefer "row group" in new prose; do not rename the option casually,
+because it is user-facing and it is in dumps.
+
+**`chunk_group_row_limit` is a different setting and does not control the group
+counters.** It is "maximum number of rows per chunk group", default 10000, from
+the 1.0-dev lineage. Setting it does not change how many groups a scan reports.
+
+**EXPLAIN's "Chunk Groups" counters count ROW GROUPS.** This is the one most
+likely to produce a wrong conclusion, so it is worth stating plainly.
+`Columnar Chunk Groups Total`, `Read`, and `Removed by Filter` are incremented
+per `NativeRowGroupMetadata` as the scan walks `rowGroupIndex`
+(`src/columnar_reader.c`, around the `groupsRead++` and `groupsSkipped++`
+lines). Measured: 200,000 rows at the default `stripe_row_limit` report 2 groups
+whatever `chunk_group_row_limit` is set to, and report 20 at
+`stripe_row_limit => 10000`. If you want a fixture with many groups to prune,
+set `stripe_row_limit`.
+
+**"Pushed-Down Filters" and "Usable Skip Predicates" are not the same number.**
+`Columnar Pushed-Down Filters` counts the scan keys the scan was HANDED.
+`Columnar Usable Skip Predicates` counts the ones the reader built a predicate
+from and can actually exclude a row group with. A filter can be pushed down and
+still exclude nothing, which is what #477 was and what #479 made visible. Read
+them together with `Chunk Groups Removed by Filter`. The second needs `ANALYZE`,
+because it describes the run rather than the plan.
+
+## Naming in the code
+
+- **`PgColumnar*`**, CamelCase, for anything with external linkage. Example:
+  `PgColumnarBeginRead`, `PgColumnarReadStats`.
+- **`pgcolumnar_*`**, lower snake case, for file-static functions, for global
+  variables backing GUCs, and for the GUCs and SQL functions themselves. Example:
+  `pgcolumnar_make_predicates`, `pgcolumnar_enable_qual_pushdown`.
+- Source lives in `src/columnar_<area>.c`. The areas worth knowing first:
+  `columnar_tableam.c` (the AM callbacks and every GUC definition),
+  `columnar_reader.c` (scan, skipping, predicates), `columnar_write_state.c`
+  (the write path and encoding selection), `columnar_customscan.c` (the scalar
+  custom scan, its costing and its EXPLAIN), `columnar_vector.c` (the two
+  vectorized aggregate nodes), `columnar_metadata.c` (the catalog).
+- Everything user-facing is in the `pgcolumnar` schema: `pgcolumnar.analyze()`,
+  `pgcolumnar.cluster()`, `pgcolumnar.set_options()`, `pgcolumnar.vacuum()`,
+  `pgcolumnar.read_parquet()`, and so on. Prefer standard SQL where standard SQL
+  exists; a function is for what SQL cannot say.
+
+## Tests
+
+A **suite** is `test/<name>.sh`. It stands up its own cluster, runs **checks**,
+and ends with `pgc_summary`.
+
+- Register every suite in `SUITES` in `test/run_all_versions.sh`. That array is
+  **one name per line and sorted**; insert in sorted position, never at the end.
+  `harness_selftest` fails if the order decays.
+- Count suites by asking the runner, never by parsing the source:
+  `bash test/run_all_versions.sh --list-suites | wc -l`. It is 132 today. A text
+  parser over the array disagrees with bash on exactly the mistake this invites,
+  and the disagreement is silent.
+- Assertions: `check` (string equality), `check_num`, `check_text` (for hashes
+  and other things `check_num` must refuse), `check_ratio`, `check_timing`.
+  `check "" ""` passes, which is why the helpers refuse empty sides.
+- **A skip is exit 66** (`PGC_EXIT_SKIPPED`). Use `pgc_skip` only for a missing
+  DEPENDENCY, which fails by default so that somebody installs the thing. For an
+  old major that lacks a core feature, `echo "SKIP ..."` then `pgc_summary` with
+  zero checks, and assert the major is a number first.
+
+## How a test is expected to argue here
+
+This is the part that distinguishes this repository, and it exists because every
+expensive mistake in its history was a check that looked like it passed rather
+than a wrong algorithm.
+
+- **Assert the premise.** Before comparing two numbers, assert the fact that
+  makes the comparison mean something: that the fixture has the rows, that the
+  planner chose the node under test, that pruning actually happened. A suite that
+  asserts a derived relation without asserting the physical fact underneath it
+  can stay green for its whole life while measuring nothing. `zonemap_cost`
+  priced pruning correctly for a year while pruning zero groups.
+- **Prove the guard by removal.** A test proves nothing until it has been seen to
+  fail with that specific guard removed. Fingerprint the installed `.so` when you
+  do it, because a shared install prefix will happily let the "before" run be the
+  fixed binary.
+- **A removal proof must fail for the STATED reason.** That a check can fail is
+  not evidence that it fails for the reason claimed. Read the failure text.
+- **Never gate on luck.** A premise that requires a random sample to miss, or
+  core to be unlucky, will fail eventually and will blame innocent code (#487).
+  Restate it as arithmetic, or print it as an observation.
+- **Pin known-wrong behaviour as an assertion, never as an echo.** Nobody reads a
+  passing suite's output. Pinned as an assertion, the fix turns the suite red and
+  forces the expectation to be updated deliberately.
+- **Never pipe a captured string into a reader whose exit status is your answer**
+  (#486). Under `pipefail` it reports "not found" whenever the writer fails.
+  Use `case` for a fixed string, or `grep -q PATTERN <<<"$var"`.
+
+## Building and running
+
+An ordinary PGXS extension: `make PG_CONFIG=/path/to/pg_config` then
+`make install`, and a suite is `bash test/<name>.sh /path/to/pg_config`. Each
+suite stands up and tears down its own cluster, so nothing needs a server
+running first. `bash test/run_all_versions.sh <pg_config> ...` runs the matrix,
+taking one `pg_config` per major as positional arguments and defaulting to
+PostgreSQL 15 through 19 when given none.
+
+- **Always pass `PG_CONFIG` to `make clean` as well as to `make`.** Objects left
+  from another major link an ABI-incompatible `.so`, and the symptom is an
+  `undefined symbol` at load time, which reads like a code defect and is not. It
+  cost an hour on 2026-08-07, where five test runs produced no output because the
+  cluster never started.
+- Build out of the source tree, or clean between majors. A suite installs into
+  the prefix its `pg_config` names, so two majors sharing a prefix will overwrite
+  each other's `.so`.
+
+Where a development environment is containerised, or PostgreSQL is not on the
+host, that is a property of the machine rather than of the project, and belongs
+in local notes beside `HANDOFF.md` rather than here.
+- Cadence: PG18 and PG19 per pull request, the full PostgreSQL 15 through 19
+  matrix per feature. The per-PR pair cannot see a version boundary below 18,
+  which is exactly how #218 was missed.
+- A red is not a result until the job's own conclusion has been read. Cancelled
+  jobs, queued jobs, and genuine failures are not distinguishable on the board.