fix(handler): run jobs without open transactions and survive lost connections (#9)

[Matys333]

Changes

• feat(output): add chunk-safe output params parser and output limiter

OutputParamsParser accumulates OUTPUT PARAMS values across output chunks and finds a line split between two chunks. OutputLimiter caps the stored output and emits a truncation marker once.

• fix(handler): run jobs without open transactions and survive lost connections

The handler kept the job entity managed and flushed it from a busy loop, rewriting the whole output on every iteration. A connection dropped inside such a flush left pdo believing in a transaction DBAL had forgotten, and the worker crashed in the retry send with "There is already an active transaction"; the message was redelivered every redeliver_timeout.

• JobOutputWriter writes output, status and result through DBAL in autocommit and resets the connection after a lost connection or a stale native transaction
• the job is claimed atomically (planned -> running), redelivered messages of running/finished jobs do not run the command again
• the wait loop sleeps 100 ms and writes at most once per poll interval; database errors are tolerated up to db_failure_tolerance
• a failed command is a job status; only an unstorable result is thrown, as UnrecoverableMessageHandlingException
• a missing job is rejected without retry
• new config job_queue.processing: poll_interval_ms, output_max_bytes, db_failure_tolerance, rerun_on_redelivery

• test(handler): add SQLite integration tests for job processing

Runs the real handler with real subprocesses against a temporary SQLite database. A DBAL driver middleware injects lost connections (including the stale pdo transaction of TT-SERVER-CQ) and the DBAL logging middleware counts statements and transactions. Adds pdo_sqlite to the CI matrix.

• docs: document job processing options and messenger recommendations

• fix(handler): store only valid UTF-8 and never let rejected output block the result

• Utf8Stream holds back a character split between two pipe reads and scrubs bytes which are not UTF-8; OutputLimiter cuts at a character boundary. A strict MySQL utf8mb4 column rejected the broken character (1366) forever.
• Output a working database rejects (not a lost connection) is replaced by a note in the loop and in finalize, so it can neither kill a healthy command nor keep the result from being stored; a rejected result write is retried without output params and status message. Stored messages are bounded.
• The writer keeps the caller's transaction nesting level for every reset, not only in the handler's finally block.
• Finalize appends its status message only below the output cap.
• Loading and claiming the job are retried (a lost claim response is verified by status + started_at) and end in a recoverable exception; a missing job is looked up again before it is rejected.
• A rerun claim appends its note in the claim UPDATE, so MySQL always reports the changed row.
• A job deleted mid-run stops its command; Process output buffers are cleared after each read; stdout and stderr are parsed separately.
• The controller cancels by a conditional UPDATE, so it never overwrites a stored result.
• Tests: handler-level TT-SERVER-CQ guard with a non-ConnectionLost driver error plus a stale native transaction (loop and finalize), multibyte cap and split reads, rejected output, caller transaction, deleted job, lost claim race, pre-claim retries.
• CI: symfony/flex so SYMFONY_REQUIRE applies, DBAL 3 / ORM 2 legs; conflict with symfony/dependency-injection 8. Docs and changelog updated.

---