This is PR 3 of the experimental postgres query stack. Adds the rest
of the input ergonomics promised in the plan and the error-formatting
polish.

Inputs: positional args become variadic, --file is repeatable, stdin
is read when neither is present, and a positional ending in '.sql'
that exists on disk is treated as a SQL file. Execution order is
files-first then positionals (cobra/pflag does not preserve interleaved
spelling, documented in --help).

Each input unit must contain exactly one statement. checkSingleStatement
walks the SQL with a hand-written conservative scanner that ignores
';' inside single-quoted strings, double-quoted identifiers, line
comments, block comments, and dollar-quoted bodies. Multi-statement
strings are rejected before connect with a hint pointing at the
multi-input alternatives.

Multi-input output:

  - text: each per-unit result rendered inline, separated by a blank
    line (mirrors psql's compact text shape).
  - json: top-level array of per-unit result objects with shape
    {"sql","kind","elapsed_ms",...}; rows-producing units carry a
    "rows":[...] array, command-only carry "command"+"rows_affected".
    Each per-unit object is buffered to completion before write; the
    outer array streams across units. The plan accepts this trade-off:
    huge SELECTs in multi-input invocations buffer.
  - csv: rejected pre-flight when N>1 (no sensible cross-schema shape).
    Single-input csv keeps streaming.

Per-unit errors render as a {"kind":"error", ...} entry in the JSON
shape so scripts can detect failure without checking exit code.
Sequential execution stops on the first failing unit; the successful
prefix is rendered.

formatPgError renders *pgconn.PgError with SEVERITY, SQLSTATE, DETAIL,
HINT inline. Non-PgError values pass through unchanged so connect-time
errors keep their original wording.

Single-input keeps the streaming sinks from PR 2; only multi-input goes
through the buffered renderer. Session state (SET, temp tables) carries
across input units because they share one connection.

TUI for >30 rows is deferred to a follow-up. The current text path uses
the static tabwriter table for both single- and multi-input.

Co-authored-by: Isaac

MUSTs:
- Multi-input JSON error envelope: thread the failing *unitResult into
  renderJSONMulti so source/sql/elapsed_ms reflect the actual failing
  input instead of empty strings.
- Canonical key order for every per-unit object:
    {"source", "sql", "kind", "elapsed_ms", payload}
  Success and error envelopes now share the same shape so consumers
  don't have to special-case kind=="error" for missing fields.

SHOULDs:
- Single-input path now goes through formatPgError, so DETAIL/HINT
  surface consistently across single- and multi-input.
- runUnitBuffered reuses executeOne via a new bufferSink. The two
  query loops collapse to one; future error-handling changes auto-
  propagate.
- Scanner: reject `$<digit>...` as a dollar-quote tag (PG docs forbid
  digit-leading tags). Pinned with a test for `SELECT $1, $2 FROM t`
  and `SELECT $1 FROM t; SELECT 2`.
- Pin the E-string over-rejection behaviour with a test, so a future
  scanner improvement has to update the assertion.

CONSIDERs:
- Capture elapsed_ms on the error path too (was previously discarded).
- Promote multiStatementHint to a const.
- Drop jsonEscapeShort (was a fragile micro-opt for an always-ASCII
  domain); use marshalJSON for the command verb instead.
- Add TestRenderJSONMulti_FirstUnitFails to pin the empty-success-
  prefix framing.

Co-authored-by: Isaac

Round-2 reviewer noted jsonErrorObject's defensive branches around
writeJSONUnitHeader/marshalJSON are unreachable (encoding/json doesn't
error on string inputs), and the repo rule says drop "just in case"
fallbacks. Replace with panic-on-impossible helpers.

Co-authored-by: Isaac

…tum/postgres-query-pr3-multi-input

# Conflicts:
#	acceptance/cmd/experimental/postgres/query/argument-errors/output.txt
#	acceptance/cmd/experimental/postgres/query/argument-errors/script

…tum/postgres-query-pr3-multi-input

# Conflicts:
#	experimental/postgres/cmd/query.go

…tum/postgres-query-pr3-multi-input

# Conflicts:
#	experimental/postgres/cmd/query.go

…tum/postgres-query-pr3-multi-input

…treaming' into simonfaltum/postgres-query-pr3-multi-input

…treaming' into simonfaltum/postgres-query-pr3-multi-input

bufferSink.Begin stashed the FieldDescription slice it was handed. pgx
reuses that slice's backing array across queries on the same connection
(pgConn.fieldDescriptions is a fixed-size buffer that's re-sliced per
statement), so each buffered unit's Fields ended up pointing at the LAST
query's row description. The multi-input renderers then emitted the wrong
column names for every unit but the last.

Clone the slice so each buffered unit owns its column descriptions.

Co-authored-by: Isaac

…treaming' into simonfaltum/postgres-query-pr3-multi-input

…treaming' into simonfaltum/postgres-query-pr3-multi-input

…treaming' into simonfaltum/postgres-query-pr3-multi-input